Upload 42 files

demo_utils.py

@@ -0,0 +1,196 @@
+# -*- coding: utf-8 -*-
+import os
+import torch
+import argparse
+import numpy as np
+import open3d as o3d
+from huggingface_hub import hf_hub_download, HfFolder
+
+from segment import seg_point, seg_box, seg_mask
+import sam2point.dataset as dataset
+import sam2point.configs as configs
+from sam2point.voxelizer import Voxelizer
+from sam2point.utils import cal
+from show import render_scene, render_scene_outdoor
+
+
+import matplotlib.pyplot as plt
+import plotly.graph_objects as go
+
+print("Torch CUDA:", torch.cuda.is_available())
+# use bfloat16 for the entire notebook
+torch.autocast(device_type="cuda", dtype=torch.bfloat16).__enter__()
+
+
+# if torch.cuda.get_device_properties(0).major >= 8:
+#    # turn on tfloat32 for Ampere GPUs (https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices)
+#    torch.backends.cuda.matmul.allow_tf32 = True
+#    torch.backends.cudnn.allow_tf32 = True
+
+
+def run_demo(dataset_name, prompt_type, sample_idx, prompt_idx, voxel_size, theta, mode, ret_prompt):
+   parser = argparse.ArgumentParser()
+   parser.add_argument('--dataset', choices=['S3DIS', 'ScanNet', 'Objaverse', 'KITTI', 'Semantic3D'], default='Objaverse', help='dataset selected')
+   parser.add_argument('--prompt_type', choices=['point', 'box', 'mask'], default='point', help='prompt type selected')
+   parser.add_argument('--sample_idx', type=int, default=2, help='the index of the scene or object')
+   parser.add_argument('--prompt_idx', type=int, default=0, help='the index of the prompt')   
+   parser.add_argument('--voxel_size', type=float, default=0.02, help='voxel size')   
+   parser.add_argument('--theta', type=float, default=0.5)  # indoor NOTE
+   parser.add_argument('--mode', type=str, default='bilinear')  # indoor NOTE
+   parser.add_argument("--ret_prompt", action="store_true")
+   args = parser.parse_args()
+   args.dataset, args.prompt_type, args.sample_idx, args.prompt_idx = dataset_name, prompt_type, sample_idx, prompt_idx
+   args.voxel_size, args.theta, args.mode, args.ret_prompt = voxel_size, theta, mode, ret_prompt
+   print(args)
+
+   #cache
+   name_list = [args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+   name = '_'.join(name_list)
+
+   # hf
+   repo_id = "ZiyuG/Cache"
+   result_name = "cache_results/" + name + '.npy'
+   prompt_name = "cache_prompt/" + name + '.npy'
+   token = os.getenv('HF_TOKEN')
+
+   try:
+       result_file = hf_hub_download(repo_id=repo_id, filename=result_name, use_auth_token=token, repo_type='dataset')
+       prompt_file = hf_hub_download(repo_id=repo_id, filename=prompt_name, use_auth_token=token, repo_type='dataset')
+       new_color = np.load(result_file)
+       PROMPT = np.load(prompt_file)
+       if not args.ret_prompt: return new_color, PROMPT
+       else:   return PROMPT
+   except Exception as e:
+       pass
+   #########
+   if args.dataset == 'S3DIS':
+       info = configs.S3DIS_samples[args.sample_idx]
+       # early return
+       if args.prompt_type == 'point' and args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       elif args.prompt_type == 'box' and args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       point, color = dataset.load_S3DIS_sample(info['path'])
+   elif args.dataset == 'ScanNet':
+       info = configs.ScanNet_samples[args.sample_idx]
+       # early return
+       if args.prompt_type == 'point' and args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       elif args.prompt_type == 'box' and args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       point, color = dataset.load_ScanNet_sample(info['path'])
+   elif args.dataset == 'Objaverse':
+       info = configs.Objaverse_samples[args.sample_idx]
+       # early return
+       if args.prompt_type == 'point' and args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       elif args.prompt_type == 'box' and args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       point, color = dataset.load_Objaverse_sample(info['path'])
+       args.voxel_size = info[configs.VOXEL[args.prompt_type]][args.prompt_idx]
+   elif args.dataset == 'KITTI':
+       info = configs.KITTI_samples[args.sample_idx]
+       # early return
+       if args.prompt_type == 'point' and args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       elif args.prompt_type == 'box' and args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       point, color = dataset.load_KITTI_sample(info['path'])
+       args.voxel_size = info[configs.VOXEL[args.prompt_type]][args.prompt_idx]
+   elif args.dataset == 'Semantic3D':
+       info = configs.Semantic3D_samples[args.sample_idx]
+       # early return
+       if args.prompt_type == 'point' and args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       elif args.prompt_type == 'box' and args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       point, color = dataset.load_Semantic3D_sample(info['path'], args.sample_idx)
+       args.voxel_size = info[configs.VOXEL[args.prompt_type]][args.prompt_idx]
+   
+   
+   point_color = np.concatenate([point, color], axis=1)
+   voxelizer = Voxelizer(voxel_size=args.voxel_size, clip_bound=None)
+  
+   labels_in = point[:, :1].astype(int)
+   locs, feats, labels, inds_reconstruct = voxelizer.voxelize(point, color, labels_in)
+
+
+   if args.prompt_type == 'point':
+       if args.ret_prompt:     return list(np.array(info['point_prompts'])[args.prompt_idx])
+       mask = seg_point(locs, feats, info['point_prompts'], args)
+       point_prompts = np.array(info['point_prompts'])
+       prompt_point = list(point_prompts[args.prompt_idx])
+       prompt_box = None
+       PROMPT = prompt_point
+   elif args.prompt_type == 'box':
+       if args.ret_prompt:     return list(np.array(info['box_prompts'])[args.prompt_idx])
+       mask = seg_box(locs, feats, info['box_prompts'], args)
+       point_prompts = np.array(info['box_prompts'])
+       prompt_point = None
+       prompt_box = list(point_prompts[args.prompt_idx])
+       PROMPT = prompt_box
+   elif args.prompt_type == 'mask':
+       if 'mask_prompts' not in info:  info['mask_prompts'] = info['point_prompts']
+       mask, prompt_mask = seg_mask(locs, feats, info['mask_prompts'], args)
+       prompt_point, prompt_box = None, None
+       point_locs = locs[inds_reconstruct]
+       point_prompt_mask = prompt_mask[point_locs[:, 0], point_locs[:, 1], point_locs[:, 2]]
+       point_prompt_mask = point_prompt_mask.unsqueeze(-1)
+       point_prompt_mask_not = ~point_prompt_mask
+       color_prompt_mask = color * point_prompt_mask_not.numpy() + (color * 0 + np.array([[1., 0., 0.]])) * point_prompt_mask.numpy()
+       PROMPT = color_prompt_mask
+       if args.ret_prompt:    
+           return color_prompt_mask
+  
+   point_locs = locs[inds_reconstruct]
+   point_mask = mask[point_locs[:, 0], point_locs[:, 1], point_locs[:, 2]]
+  
+   point_mask = point_mask.unsqueeze(-1)
+   point_mask_not = ~point_mask
+  
+   point, color = point_color[:, :3], point_color[:, 3:]
+   new_color = color * point_mask_not.numpy() + (color * 0 + np.array([[0., 1., 0.]])) * point_mask.numpy()
+
+   name_list = [args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+   name = '_'.join(name_list) + 'frames'
+   # os.system('rm -rf ' + name)
+
+   #cache
+   name_list = [args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+   name = '_'.join(name_list)
+   np.save("./cache_results/" + name + '.npy', new_color)
+   np.save("./cache_prompt/" + name + '.npy', PROMPT)
+   return new_color, PROMPT
+
+
+
+
+def create_box(prompt):
+   x_min, y_min, z_min, x_max, y_max, z_max = tuple(prompt)
+   bbox_points = np.array([
+       [x_min, y_min, z_min],
+       [x_max, y_min, z_min],
+       [x_max, y_max, z_min],
+       [x_min, y_max, z_min],
+       [x_min, y_min, z_max],
+       [x_max, y_min, z_max],
+       [x_max, y_max, z_max],
+       [x_min, y_max, z_max]
+   ])
+
+
+   edges = [
+       (0, 1), (1, 2), (2, 3), (3, 0), # Bottom face
+       (4, 5), (5, 6), (6, 7), (7, 4), # Top face
+       (0, 4), (1, 5), (2, 6), (3, 7)  # Vertical edges
+   ]
+
+
+   bbox_lines = []
+   f = 1
+   for start, end in edges:
+       bbox_lines.append(go.Scatter3d(
+           x=[bbox_points[start, 0], bbox_points[end, 0]],
+           y=[bbox_points[start, 1], bbox_points[end, 1]],
+           z=[bbox_points[start, 2], bbox_points[end, 2]],
+           mode='lines',
+        #    line=dict(color='red', width=2),  # Customize color and width
+        #    line=dict(color='rgb(255, 140, 0)', width=4),  # Customize color and width
+           line=dict(color='rgb(220, 20, 60)', width=6),  # Customize color and width
+           name="Box Prompt" if f == 1 else "",
+           showlegend=True if f == 1 else False
+       ))
+       f = 0
+   return bbox_lines
+
+

sam2/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from hydra import initialize_config_module
+
+initialize_config_module("sam2_configs", version_base="1.2")

sam2/automatic_mask_generator.py

@@ -0,0 +1,434 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/automatic_mask_generator.py
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from torchvision.ops.boxes import batched_nms, box_area  # type: ignore
+
+from sam2.modeling.sam2_base import SAM2Base
+from sam2.sam2_image_predictor import SAM2ImagePredictor
+from sam2.utils.amg import (
+    area_from_rle,
+    batch_iterator,
+    batched_mask_to_box,
+    box_xyxy_to_xywh,
+    build_all_layer_point_grids,
+    calculate_stability_score,
+    coco_encode_rle,
+    generate_crop_boxes,
+    is_box_near_crop_edge,
+    mask_to_rle_pytorch,
+    MaskData,
+    remove_small_regions,
+    rle_to_mask,
+    uncrop_boxes_xyxy,
+    uncrop_masks,
+    uncrop_points,
+)
+
+
+class SAM2AutomaticMaskGenerator:
+    def __init__(
+        self,
+        model: SAM2Base,
+        points_per_side: Optional[int] = 32,
+        points_per_batch: int = 64,
+        pred_iou_thresh: float = 0.8,
+        stability_score_thresh: float = 0.95,
+        stability_score_offset: float = 1.0,
+        mask_threshold: float = 0.0,
+        box_nms_thresh: float = 0.7,
+        crop_n_layers: int = 0,
+        crop_nms_thresh: float = 0.7,
+        crop_overlap_ratio: float = 512 / 1500,
+        crop_n_points_downscale_factor: int = 1,
+        point_grids: Optional[List[np.ndarray]] = None,
+        min_mask_region_area: int = 0,
+        output_mode: str = "binary_mask",
+        use_m2m: bool = False,
+        multimask_output: bool = True,
+    ) -> None:
+        """
+        Using a SAM 2 model, generates masks for the entire image.
+        Generates a grid of point prompts over the image, then filters
+        low quality and duplicate masks. The default settings are chosen
+        for SAM 2 with a HieraL backbone.
+
+        Arguments:
+          model (Sam): The SAM 2 model to use for mask prediction.
+          points_per_side (int or None): The number of points to be sampled
+            along one side of the image. The total number of points is
+            points_per_side**2. If None, 'point_grids' must provide explicit
+            point sampling.
+          points_per_batch (int): Sets the number of points run simultaneously
+            by the model. Higher numbers may be faster but use more GPU memory.
+          pred_iou_thresh (float): A filtering threshold in [0,1], using the
+            model's predicted mask quality.
+          stability_score_thresh (float): A filtering threshold in [0,1], using
+            the stability of the mask under changes to the cutoff used to binarize
+            the model's mask predictions.
+          stability_score_offset (float): The amount to shift the cutoff when
+            calculated the stability score.
+          mask_threshold (float): Threshold for binarizing the mask logits
+          box_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks.
+          crop_n_layers (int): If >0, mask prediction will be run again on
+            crops of the image. Sets the number of layers to run, where each
+            layer has 2**i_layer number of image crops.
+          crop_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks between different crops.
+          crop_overlap_ratio (float): Sets the degree to which crops overlap.
+            In the first crop layer, crops will overlap by this fraction of
+            the image length. Later layers with more crops scale down this overlap.
+          crop_n_points_downscale_factor (int): The number of points-per-side
+            sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
+          point_grids (list(np.ndarray) or None): A list over explicit grids
+            of points used for sampling, normalized to [0,1]. The nth grid in the
+            list is used in the nth crop layer. Exclusive with points_per_side.
+          min_mask_region_area (int): If >0, postprocessing will be applied
+            to remove disconnected regions and holes in masks with area smaller
+            than min_mask_region_area. Requires opencv.
+          output_mode (str): The form masks are returned in. Can be 'binary_mask',
+            'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
+            For large resolutions, 'binary_mask' may consume large amounts of
+            memory.
+          use_m2m (bool): Whether to add a one step refinement using previous mask predictions.
+          multimask_output (bool): Whether to output multimask at each point of the grid.
+        """
+
+        assert (points_per_side is None) != (
+            point_grids is None
+        ), "Exactly one of points_per_side or point_grid must be provided."
+        if points_per_side is not None:
+            self.point_grids = build_all_layer_point_grids(
+                points_per_side,
+                crop_n_layers,
+                crop_n_points_downscale_factor,
+            )
+        elif point_grids is not None:
+            self.point_grids = point_grids
+        else:
+            raise ValueError("Can't have both points_per_side and point_grid be None.")
+
+        assert output_mode in [
+            "binary_mask",
+            "uncompressed_rle",
+            "coco_rle",
+        ], f"Unknown output_mode {output_mode}."
+        if output_mode == "coco_rle":
+            try:
+                from pycocotools import mask as mask_utils  # type: ignore  # noqa: F401
+            except ImportError as e:
+                print("Please install pycocotools")
+                raise e
+
+        self.predictor = SAM2ImagePredictor(
+            model,
+            max_hole_area=min_mask_region_area,
+            max_sprinkle_area=min_mask_region_area,
+        )
+        self.points_per_batch = points_per_batch
+        self.pred_iou_thresh = pred_iou_thresh
+        self.stability_score_thresh = stability_score_thresh
+        self.stability_score_offset = stability_score_offset
+        self.mask_threshold = mask_threshold
+        self.box_nms_thresh = box_nms_thresh
+        self.crop_n_layers = crop_n_layers
+        self.crop_nms_thresh = crop_nms_thresh
+        self.crop_overlap_ratio = crop_overlap_ratio
+        self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
+        self.min_mask_region_area = min_mask_region_area
+        self.output_mode = output_mode
+        self.use_m2m = use_m2m
+        self.multimask_output = multimask_output
+
+    @torch.no_grad()
+    def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
+        """
+        Generates masks for the given image.
+
+        Arguments:
+          image (np.ndarray): The image to generate masks for, in HWC uint8 format.
+
+        Returns:
+           list(dict(str, any)): A list over records for masks. Each record is
+             a dict containing the following keys:
+               segmentation (dict(str, any) or np.ndarray): The mask. If
+                 output_mode='binary_mask', is an array of shape HW. Otherwise,
+                 is a dictionary containing the RLE.
+               bbox (list(float)): The box around the mask, in XYWH format.
+               area (int): The area in pixels of the mask.
+               predicted_iou (float): The model's own prediction of the mask's
+                 quality. This is filtered by the pred_iou_thresh parameter.
+               point_coords (list(list(float))): The point coordinates input
+                 to the model to generate this mask.
+               stability_score (float): A measure of the mask's quality. This
+                 is filtered on using the stability_score_thresh parameter.
+               crop_box (list(float)): The crop of the image used to generate
+                 the mask, given in XYWH format.
+        """
+
+        # Generate masks
+        mask_data = self._generate_masks(image)
+
+        # Encode masks
+        if self.output_mode == "coco_rle":
+            mask_data["segmentations"] = [
+                coco_encode_rle(rle) for rle in mask_data["rles"]
+            ]
+        elif self.output_mode == "binary_mask":
+            mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
+        else:
+            mask_data["segmentations"] = mask_data["rles"]
+
+        # Write mask records
+        curr_anns = []
+        for idx in range(len(mask_data["segmentations"])):
+            ann = {
+                "segmentation": mask_data["segmentations"][idx],
+                "area": area_from_rle(mask_data["rles"][idx]),
+                "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
+                "predicted_iou": mask_data["iou_preds"][idx].item(),
+                "point_coords": [mask_data["points"][idx].tolist()],
+                "stability_score": mask_data["stability_score"][idx].item(),
+                "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
+            }
+            curr_anns.append(ann)
+
+        return curr_anns
+
+    def _generate_masks(self, image: np.ndarray) -> MaskData:
+        orig_size = image.shape[:2]
+        crop_boxes, layer_idxs = generate_crop_boxes(
+            orig_size, self.crop_n_layers, self.crop_overlap_ratio
+        )
+
+        # Iterate over image crops
+        data = MaskData()
+        for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
+            crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
+            data.cat(crop_data)
+
+        # Remove duplicate masks between crops
+        if len(crop_boxes) > 1:
+            # Prefer masks from smaller crops
+            scores = 1 / box_area(data["crop_boxes"])
+            scores = scores.to(data["boxes"].device)
+            keep_by_nms = batched_nms(
+                data["boxes"].float(),
+                scores,
+                torch.zeros_like(data["boxes"][:, 0]),  # categories
+                iou_threshold=self.crop_nms_thresh,
+            )
+            data.filter(keep_by_nms)
+        data.to_numpy()
+        return data
+
+    def _process_crop(
+        self,
+        image: np.ndarray,
+        crop_box: List[int],
+        crop_layer_idx: int,
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        # Crop the image and calculate embeddings
+        x0, y0, x1, y1 = crop_box
+        cropped_im = image[y0:y1, x0:x1, :]
+        cropped_im_size = cropped_im.shape[:2]
+        self.predictor.set_image(cropped_im)
+
+        # Get points for this crop
+        points_scale = np.array(cropped_im_size)[None, ::-1]
+        points_for_image = self.point_grids[crop_layer_idx] * points_scale
+
+        # Generate masks for this crop in batches
+        data = MaskData()
+        for (points,) in batch_iterator(self.points_per_batch, points_for_image):
+            batch_data = self._process_batch(
+                points, cropped_im_size, crop_box, orig_size, normalize=True
+            )
+            data.cat(batch_data)
+            del batch_data
+        self.predictor.reset_predictor()
+
+        # Remove duplicates within this crop.
+        keep_by_nms = batched_nms(
+            data["boxes"].float(),
+            data["iou_preds"],
+            torch.zeros_like(data["boxes"][:, 0]),  # categories
+            iou_threshold=self.box_nms_thresh,
+        )
+        data.filter(keep_by_nms)
+
+        # Return to the original image frame
+        data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
+        data["points"] = uncrop_points(data["points"], crop_box)
+        data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
+
+        return data
+
+    def _process_batch(
+        self,
+        points: np.ndarray,
+        im_size: Tuple[int, ...],
+        crop_box: List[int],
+        orig_size: Tuple[int, ...],
+        normalize=False,
+    ) -> MaskData:
+        orig_h, orig_w = orig_size
+
+        # Run model on this batch
+        points = torch.as_tensor(points, device=self.predictor.device)
+        in_points = self.predictor._transforms.transform_coords(
+            points, normalize=normalize, orig_hw=im_size
+        )
+        in_labels = torch.ones(
+            in_points.shape[0], dtype=torch.int, device=in_points.device
+        )
+        masks, iou_preds, low_res_masks = self.predictor._predict(
+            in_points[:, None, :],
+            in_labels[:, None],
+            multimask_output=self.multimask_output,
+            return_logits=True,
+        )
+
+        # Serialize predictions and store in MaskData
+        data = MaskData(
+            masks=masks.flatten(0, 1),
+            iou_preds=iou_preds.flatten(0, 1),
+            points=points.repeat_interleave(masks.shape[1], dim=0),
+            low_res_masks=low_res_masks.flatten(0, 1),
+        )
+        del masks
+
+        if not self.use_m2m:
+            # Filter by predicted IoU
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            # Calculate and filter by stability score
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+        else:
+            # One step refinement using previous mask predictions
+            in_points = self.predictor._transforms.transform_coords(
+                data["points"], normalize=normalize, orig_hw=im_size
+            )
+            labels = torch.ones(
+                in_points.shape[0], dtype=torch.int, device=in_points.device
+            )
+            masks, ious = self.refine_with_m2m(
+                in_points, labels, data["low_res_masks"], self.points_per_batch
+            )
+            data["masks"] = masks.squeeze(1)
+            data["iou_preds"] = ious.squeeze(1)
+
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+
+        # Threshold masks and calculate boxes
+        data["masks"] = data["masks"] > self.mask_threshold
+        data["boxes"] = batched_mask_to_box(data["masks"])
+
+        # Filter boxes that touch crop boundaries
+        keep_mask = ~is_box_near_crop_edge(
+            data["boxes"], crop_box, [0, 0, orig_w, orig_h]
+        )
+        if not torch.all(keep_mask):
+            data.filter(keep_mask)
+
+        # Compress to RLE
+        data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
+        data["rles"] = mask_to_rle_pytorch(data["masks"])
+        del data["masks"]
+
+        return data
+
+    @staticmethod
+    def postprocess_small_regions(
+        mask_data: MaskData, min_area: int, nms_thresh: float
+    ) -> MaskData:
+        """
+        Removes small disconnected regions and holes in masks, then reruns
+        box NMS to remove any new duplicates.
+
+        Edits mask_data in place.
+
+        Requires open-cv as a dependency.
+        """
+        if len(mask_data["rles"]) == 0:
+            return mask_data
+
+        # Filter small disconnected regions and holes
+        new_masks = []
+        scores = []
+        for rle in mask_data["rles"]:
+            mask = rle_to_mask(rle)
+
+            mask, changed = remove_small_regions(mask, min_area, mode="holes")
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask, min_area, mode="islands")
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores),
+            torch.zeros_like(boxes[:, 0]),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
+                mask_data["boxes"][i_mask] = boxes[i_mask]  # update res directly
+        mask_data.filter(keep_by_nms)
+
+        return mask_data
+
+    def refine_with_m2m(self, points, point_labels, low_res_masks, points_per_batch):
+        new_masks = []
+        new_iou_preds = []
+
+        for cur_points, cur_point_labels, low_res_mask in batch_iterator(
+            points_per_batch, points, point_labels, low_res_masks
+        ):
+            best_masks, best_iou_preds, _ = self.predictor._predict(
+                cur_points[:, None, :],
+                cur_point_labels[:, None],
+                mask_input=low_res_mask[:, None, :],
+                multimask_output=False,
+                return_logits=True,
+            )
+            new_masks.append(best_masks)
+            new_iou_preds.append(best_iou_preds)
+        masks = torch.cat(new_masks, dim=0)
+        return masks, torch.cat(new_iou_preds, dim=0)

sam2/build_sam.py

@@ -0,0 +1,127 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from hydra import compose
+from hydra.utils import instantiate
+from omegaconf import OmegaConf
+
+
+def build_sam2(
+    config_file,
+    ckpt_path=None,
+    device="cuda",
+    mode="eval",
+    hydra_overrides_extra=[],
+    apply_postprocessing=True,
+):
+
+    if apply_postprocessing:
+        hydra_overrides_extra = hydra_overrides_extra.copy()
+        hydra_overrides_extra += [
+            # dynamically fall back to multi-mask if the single mask is not stable
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_via_stability=true",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_delta=0.05",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_thresh=0.98",
+        ]
+    # Read config and init model
+    cfg = compose(config_name=config_file, overrides=hydra_overrides_extra)
+    OmegaConf.resolve(cfg)
+    model = instantiate(cfg.model, _recursive_=True)
+    _load_checkpoint(model, ckpt_path)
+    model = model.to(device)
+    if mode == "eval":
+        model.eval()
+    return model
+
+
+def build_sam2_video_predictor(
+    config_file,
+    ckpt_path=None,
+    device="cuda",
+    mode="eval",
+    hydra_overrides_extra=[],
+    apply_postprocessing=True,
+):
+    hydra_overrides = [
+        "++model._target_=sam2.sam2_video_predictor.SAM2VideoPredictor",
+    ]
+    if apply_postprocessing:
+        hydra_overrides_extra = hydra_overrides_extra.copy()
+        hydra_overrides_extra += [
+            # dynamically fall back to multi-mask if the single mask is not stable
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_via_stability=true",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_delta=0.05",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_thresh=0.98",
+            # the sigmoid mask logits on interacted frames with clicks in the memory encoder so that the encoded masks are exactly as what users see from clicking
+            "++model.binarize_mask_from_pts_for_mem_enc=true",
+            # fill small holes in the low-res masks up to `fill_hole_area` (before resizing them to the original video resolution)
+            "++model.fill_hole_area=8",
+        ]
+    hydra_overrides.extend(hydra_overrides_extra)
+
+    # Read config and init model
+    cfg = compose(config_name=config_file, overrides=hydra_overrides)
+    OmegaConf.resolve(cfg)
+    model = instantiate(cfg.model, _recursive_=True)
+    _load_checkpoint(model, ckpt_path)
+    model = model.to(device)
+    if mode == "eval":
+        model.eval()
+    return model
+
+
+def build_sam2_hf(model_id, **kwargs):
+
+    from huggingface_hub import hf_hub_download
+
+    model_id_to_filenames = {
+        "facebook/sam2-hiera-tiny": ("sam2_hiera_t.yaml", "sam2_hiera_tiny.pt"),
+        "facebook/sam2-hiera-small": ("sam2_hiera_s.yaml", "sam2_hiera_small.pt"),
+        "facebook/sam2-hiera-base-plus": (
+            "sam2_hiera_b+.yaml",
+            "sam2_hiera_base_plus.pt",
+        ),
+        "facebook/sam2-hiera-large": ("sam2_hiera_l.yaml", "sam2_hiera_large.pt"),
+    }
+    config_name, checkpoint_name = model_id_to_filenames[model_id]
+    ckpt_path = hf_hub_download(repo_id=model_id, filename=checkpoint_name)
+    return build_sam2(config_file=config_name, ckpt_path=ckpt_path, **kwargs)
+
+
+def build_sam2_video_predictor_hf(model_id, **kwargs):
+
+    from huggingface_hub import hf_hub_download
+
+    model_id_to_filenames = {
+        "facebook/sam2-hiera-tiny": ("sam2_hiera_t.yaml", "sam2_hiera_tiny.pt"),
+        "facebook/sam2-hiera-small": ("sam2_hiera_s.yaml", "sam2_hiera_small.pt"),
+        "facebook/sam2-hiera-base-plus": (
+            "sam2_hiera_b+.yaml",
+            "sam2_hiera_base_plus.pt",
+        ),
+        "facebook/sam2-hiera-large": ("sam2_hiera_l.yaml", "sam2_hiera_large.pt"),
+    }
+    config_name, checkpoint_name = model_id_to_filenames[model_id]
+    ckpt_path = hf_hub_download(repo_id=model_id, filename=checkpoint_name)
+    return build_sam2_video_predictor(
+        config_file=config_name, ckpt_path=ckpt_path, **kwargs
+    )
+
+
+def _load_checkpoint(model, ckpt_path):
+    if ckpt_path is not None:
+        sd = torch.load(ckpt_path, map_location="cpu")["model"]
+        missing_keys, unexpected_keys = model.load_state_dict(sd)
+        if missing_keys:
+            logging.error(missing_keys)
+            raise RuntimeError()
+        if unexpected_keys:
+            logging.error(unexpected_keys)
+            raise RuntimeError()
+        logging.info("Loaded checkpoint sucessfully")

sam2/csrc/__init__.py

@@ -0,0 +1 @@
+from sam2.csrc.connected_comp import connect

sam2/csrc/backend.py

@@ -0,0 +1,12 @@
+import os
+
+from torch.utils.cpp_extension import load
+
+_src_path = os.path.dirname(os.path.abspath(__file__))
+_backend = load(name='_pvcnn_backend',
+                extra_cflags=['-O3', '-std=c++17'],
+                #extra_cuda_cflags=['--compiler-bindir=/usr/bin/gcc-8'],
+                sources=[os.path.join(_src_path, f) for f in ['connected_components.cu']]
+                )
+
+__all__ = ['_backend']

sam2/csrc/connected_comp.py

@@ -0,0 +1,9 @@
+from torch.autograd import Function
+
+from sam2.csrc.backend import _backend
+
+__all__ = ['connect']
+
+
+def connect(mask):
+    return _backend.get_connected_componnets(mask)

sam2/csrc/connected_components.cu

@@ -0,0 +1,289 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+
+// This source code is licensed under the license found in the
+// LICENSE file in the root directory of this source tree.
+
+// adapted from https://github.com/zsef123/Connected_components_PyTorch
+// with license found in the LICENSE_cctorch file in the root directory.
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <torch/script.h>
+#include <vector>
+
+// 2d
+#define BLOCK_ROWS 16
+#define BLOCK_COLS 16
+
+namespace cc2d {
+
+template <typename T>
+__device__ __forceinline__ unsigned char hasBit(T bitmap, unsigned char pos) {
+  return (bitmap >> pos) & 1;
+}
+
+__device__ int32_t find(const int32_t* s_buf, int32_t n) {
+  while (s_buf[n] != n)
+    n = s_buf[n];
+  return n;
+}
+
+__device__ int32_t find_n_compress(int32_t* s_buf, int32_t n) {
+  const int32_t id = n;
+  while (s_buf[n] != n) {
+    n = s_buf[n];
+    s_buf[id] = n;
+  }
+  return n;
+}
+
+__device__ void union_(int32_t* s_buf, int32_t a, int32_t b) {
+  bool done;
+  do {
+    a = find(s_buf, a);
+    b = find(s_buf, b);
+
+    if (a < b) {
+      int32_t old = atomicMin(s_buf + b, a);
+      done = (old == b);
+      b = old;
+    } else if (b < a) {
+      int32_t old = atomicMin(s_buf + a, b);
+      done = (old == a);
+      a = old;
+    } else
+      done = true;
+
+  } while (!done);
+}
+
+__global__ void
+init_labeling(int32_t* label, const uint32_t W, const uint32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row < H && col < W)
+    label[idx] = idx;
+}
+
+__global__ void
+merge(uint8_t* img, int32_t* label, const uint32_t W, const uint32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  uint32_t P = 0;
+
+  if (img[idx])
+    P |= 0x777;
+  if (row + 1 < H && img[idx + W])
+    P |= 0x777 << 4;
+  if (col + 1 < W && img[idx + 1])
+    P |= 0x777 << 1;
+
+  if (col == 0)
+    P &= 0xEEEE;
+  if (col + 1 >= W)
+    P &= 0x3333;
+  else if (col + 2 >= W)
+    P &= 0x7777;
+
+  if (row == 0)
+    P &= 0xFFF0;
+  if (row + 1 >= H)
+    P &= 0xFF;
+
+  if (P > 0) {
+    // If need check about top-left pixel(if flag the first bit) and hit the
+    // top-left pixel
+    if (hasBit(P, 0) && img[idx - W - 1]) {
+      union_(label, idx, idx - 2 * W - 2); // top left block
+    }
+
+    if ((hasBit(P, 1) && img[idx - W]) || (hasBit(P, 2) && img[idx - W + 1]))
+      union_(label, idx, idx - 2 * W); // top bottom block
+
+    if (hasBit(P, 3) && img[idx + 2 - W])
+      union_(label, idx, idx - 2 * W + 2); // top right block
+
+    if ((hasBit(P, 4) && img[idx - 1]) || (hasBit(P, 8) && img[idx + W - 1]))
+      union_(label, idx, idx - 2); // just left block
+  }
+}
+
+__global__ void compression(int32_t* label, const int32_t W, const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row < H && col < W)
+    find_n_compress(label, idx);
+}
+
+__global__ void final_labeling(
+    const uint8_t* img,
+    int32_t* label,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx] + 1;
+
+  if (img[idx])
+    label[idx] = y;
+  else
+    label[idx] = 0;
+
+  if (col + 1 < W) {
+    if (img[idx + 1])
+      label[idx + 1] = y;
+    else
+      label[idx + 1] = 0;
+
+    if (row + 1 < H) {
+      if (img[idx + W + 1])
+        label[idx + W + 1] = y;
+      else
+        label[idx + W + 1] = 0;
+    }
+  }
+
+  if (row + 1 < H) {
+    if (img[idx + W])
+      label[idx + W] = y;
+    else
+      label[idx + W] = 0;
+  }
+}
+
+__global__ void init_counting(
+    const int32_t* label,
+    int32_t* count_init,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx];
+  if (y > 0) {
+    int32_t count_idx = y - 1;
+    atomicAdd(count_init + count_idx, 1);
+  }
+}
+
+__global__ void final_counting(
+    const int32_t* label,
+    const int32_t* count_init,
+    int32_t* count_final,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx];
+  if (y > 0) {
+    int32_t count_idx = y - 1;
+    count_final[idx] = count_init[count_idx];
+  } else {
+    count_final[idx] = 0;
+  }
+}
+
+} // namespace cc2d
+
+std::vector<torch::Tensor> get_connected_componnets(
+    const torch::Tensor& inputs) {
+  AT_ASSERTM(inputs.is_cuda(), "inputs must be a CUDA tensor");
+  AT_ASSERTM(inputs.ndimension() == 4, "inputs must be [N, 1, H, W] shape");
+  AT_ASSERTM(
+      inputs.scalar_type() == torch::kUInt8, "inputs must be a uint8 type");
+
+  const uint32_t N = inputs.size(0);
+  const uint32_t C = inputs.size(1);
+  const uint32_t H = inputs.size(2);
+  const uint32_t W = inputs.size(3);
+
+  AT_ASSERTM(C == 1, "inputs must be [N, 1, H, W] shape");
+  AT_ASSERTM((H % 2) == 0, "height must be an even number");
+  AT_ASSERTM((W % 2) == 0, "width must be an even number");
+
+  // label must be uint32_t
+  auto label_options =
+      torch::TensorOptions().dtype(torch::kInt32).device(inputs.device());
+  torch::Tensor labels = torch::zeros({N, C, H, W}, label_options);
+  torch::Tensor counts_init = torch::zeros({N, C, H, W}, label_options);
+  torch::Tensor counts_final = torch::zeros({N, C, H, W}, label_options);
+
+  dim3 grid = dim3(
+      ((W + 1) / 2 + BLOCK_COLS - 1) / BLOCK_COLS,
+      ((H + 1) / 2 + BLOCK_ROWS - 1) / BLOCK_ROWS);
+  dim3 block = dim3(BLOCK_COLS, BLOCK_ROWS);
+  dim3 grid_count =
+      dim3((W + BLOCK_COLS) / BLOCK_COLS, (H + BLOCK_ROWS) / BLOCK_ROWS);
+  dim3 block_count = dim3(BLOCK_COLS, BLOCK_ROWS);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  for (int n = 0; n < N; n++) {
+    uint32_t offset = n * H * W;
+
+    cc2d::init_labeling<<<grid, block, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset, W, H);
+    cc2d::merge<<<grid, block, 0, stream>>>(
+        inputs.data_ptr<uint8_t>() + offset,
+        labels.data_ptr<int32_t>() + offset,
+        W,
+        H);
+    cc2d::compression<<<grid, block, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset, W, H);
+    cc2d::final_labeling<<<grid, block, 0, stream>>>(
+        inputs.data_ptr<uint8_t>() + offset,
+        labels.data_ptr<int32_t>() + offset,
+        W,
+        H);
+
+    // get the counting of each pixel
+    cc2d::init_counting<<<grid_count, block_count, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset,
+        counts_init.data_ptr<int32_t>() + offset,
+        W,
+        H);
+    cc2d::final_counting<<<grid_count, block_count, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset,
+        counts_init.data_ptr<int32_t>() + offset,
+        counts_final.data_ptr<int32_t>() + offset,
+        W,
+        H);
+  }
+
+  // returned values are [labels, counts]
+  std::vector<torch::Tensor> outputs;
+  outputs.push_back(labels);
+  outputs.push_back(counts_final);
+  return outputs;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+      "get_connected_componnets",
+      &get_connected_componnets,
+      "get_connected_componnets");
+}

sam2/modeling/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam2/modeling/backbones/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam2/modeling/backbones/hieradet.py

@@ -0,0 +1,291 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+from typing import List, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from sam2.modeling.backbones.utils import (
+    PatchEmbed,
+    window_partition,
+    window_unpartition,
+)
+
+from sam2.modeling.sam2_utils import DropPath, MLP
+
+
+def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
+    if pool is None:
+        return x
+    # (B, H, W, C) -> (B, C, H, W)
+    x = x.permute(0, 3, 1, 2)
+    x = pool(x)
+    # (B, C, H', W') -> (B, H', W', C)
+    x = x.permute(0, 2, 3, 1)
+    if norm:
+        x = norm(x)
+
+    return x
+
+
+class MultiScaleAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        q_pool: nn.Module = None,
+    ):
+        super().__init__()
+
+        self.dim = dim
+        self.dim_out = dim_out
+        self.num_heads = num_heads
+        self.q_pool = q_pool
+        self.qkv = nn.Linear(dim, dim_out * 3)
+        self.proj = nn.Linear(dim_out, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (B, H * W, 3, nHead, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
+        # q, k, v with shape (B, H * W, nheads, C)
+        q, k, v = torch.unbind(qkv, 2)
+
+        # Q pooling (for downsample at stage changes)
+        if self.q_pool:
+            q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
+            H, W = q.shape[1:3]  # downsampled shape
+            q = q.reshape(B, H * W, self.num_heads, -1)
+
+        # Torch's SDPA expects [B, nheads, H*W, C] so we transpose
+        x = F.scaled_dot_product_attention(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+        )
+        # Transpose back
+        x = x.transpose(1, 2)
+        x = x.reshape(B, H, W, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+class MultiScaleBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        drop_path: float = 0.0,
+        norm_layer: Union[nn.Module, str] = "LayerNorm",
+        q_stride: Tuple[int, int] = None,
+        act_layer: nn.Module = nn.GELU,
+        window_size: int = 0,
+    ):
+        super().__init__()
+
+        if isinstance(norm_layer, str):
+            norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)
+
+        self.dim = dim
+        self.dim_out = dim_out
+        self.norm1 = norm_layer(dim)
+
+        self.window_size = window_size
+
+        self.pool, self.q_stride = None, q_stride
+        if self.q_stride:
+            self.pool = nn.MaxPool2d(
+                kernel_size=q_stride, stride=q_stride, ceil_mode=False
+            )
+
+        self.attn = MultiScaleAttention(
+            dim,
+            dim_out,
+            num_heads=num_heads,
+            q_pool=self.pool,
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim_out)
+        self.mlp = MLP(
+            dim_out,
+            int(dim_out * mlp_ratio),
+            dim_out,
+            num_layers=2,
+            activation=act_layer,
+        )
+
+        if dim != dim_out:
+            self.proj = nn.Linear(dim, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x  # B, H, W, C
+        x = self.norm1(x)
+
+        # Skip connection
+        if self.dim != self.dim_out:
+            shortcut = do_pool(self.proj(x), self.pool)
+
+        # Window partition
+        window_size = self.window_size
+        if window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, window_size)
+
+        # Window Attention + Q Pooling (if stage change)
+        x = self.attn(x)
+        if self.q_stride:
+            # Shapes have changed due to Q pooling
+            window_size = self.window_size // self.q_stride[0]
+            H, W = shortcut.shape[1:3]
+
+            pad_h = (window_size - H % window_size) % window_size
+            pad_w = (window_size - W % window_size) % window_size
+            pad_hw = (H + pad_h, W + pad_w)
+
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, window_size, pad_hw, (H, W))
+
+        x = shortcut + self.drop_path(x)
+        # MLP
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class Hiera(nn.Module):
+    """
+    Reference: https://arxiv.org/abs/2306.00989
+    """
+
+    def __init__(
+        self,
+        embed_dim: int = 96,  # initial embed dim
+        num_heads: int = 1,  # initial number of heads
+        drop_path_rate: float = 0.0,  # stochastic depth
+        q_pool: int = 3,  # number of q_pool stages
+        q_stride: Tuple[int, int] = (2, 2),  # downsample stride bet. stages
+        stages: Tuple[int, ...] = (2, 3, 16, 3),  # blocks per stage
+        dim_mul: float = 2.0,  # dim_mul factor at stage shift
+        head_mul: float = 2.0,  # head_mul factor at stage shift
+        window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
+        # window size per stage, when not using global att.
+        window_spec: Tuple[int, ...] = (
+            8,
+            4,
+            14,
+            7,
+        ),
+        # global attn in these blocks
+        global_att_blocks: Tuple[int, ...] = (
+            12,
+            16,
+            20,
+        ),
+        return_interm_layers=True,  # return feats from every stage
+    ):
+        super().__init__()
+
+        assert len(stages) == len(window_spec)
+        self.window_spec = window_spec
+
+        depth = sum(stages)
+        self.q_stride = q_stride
+        self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
+        assert 0 <= q_pool <= len(self.stage_ends[:-1])
+        self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
+        self.return_interm_layers = return_interm_layers
+
+        self.patch_embed = PatchEmbed(
+            embed_dim=embed_dim,
+        )
+        # Which blocks have global att?
+        self.global_att_blocks = global_att_blocks
+
+        # Windowed positional embedding (https://arxiv.org/abs/2311.05613)
+        self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
+        self.pos_embed = nn.Parameter(
+            torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
+        )
+        self.pos_embed_window = nn.Parameter(
+            torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
+        )
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+
+        cur_stage = 1
+        self.blocks = nn.ModuleList()
+
+        for i in range(depth):
+            dim_out = embed_dim
+            # lags by a block, so first block of
+            # next stage uses an initial window size
+            # of previous stage and final window size of current stage
+            window_size = self.window_spec[cur_stage - 1]
+
+            if self.global_att_blocks is not None:
+                window_size = 0 if i in self.global_att_blocks else window_size
+
+            if i - 1 in self.stage_ends:
+                dim_out = int(embed_dim * dim_mul)
+                num_heads = int(num_heads * head_mul)
+                cur_stage += 1
+
+            block = MultiScaleBlock(
+                dim=embed_dim,
+                dim_out=dim_out,
+                num_heads=num_heads,
+                drop_path=dpr[i],
+                q_stride=self.q_stride if i in self.q_pool_blocks else None,
+                window_size=window_size,
+            )
+
+            embed_dim = dim_out
+            self.blocks.append(block)
+
+        self.channel_list = (
+            [self.blocks[i].dim_out for i in self.stage_ends[::-1]]
+            if return_interm_layers
+            else [self.blocks[-1].dim_out]
+        )
+
+    def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
+        h, w = hw
+        window_embed = self.pos_embed_window
+        pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
+        pos_embed = pos_embed + window_embed.tile(
+            [x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
+        )
+        pos_embed = pos_embed.permute(0, 2, 3, 1)
+        return pos_embed
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.patch_embed(x)
+        # x: (B, H, W, C)
+
+        # Add pos embed
+        x = x + self._get_pos_embed(x.shape[1:3])
+
+        outputs = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if (i == self.stage_ends[-1]) or (
+                i in self.stage_ends and self.return_interm_layers
+            ):
+                feats = x.permute(0, 3, 1, 2)
+                outputs.append(feats)
+
+        return outputs

sam2/modeling/backbones/image_encoder.py

@@ -0,0 +1,133 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class ImageEncoder(nn.Module):
+    def __init__(
+        self,
+        trunk: nn.Module,
+        neck: nn.Module,
+        scalp: int = 0,
+    ):
+        super().__init__()
+        self.trunk = trunk
+        self.neck = neck
+        self.scalp = scalp
+        assert (
+            self.trunk.channel_list == self.neck.backbone_channel_list
+        ), f"Channel dims of trunk and neck do not match. Trunk: {self.trunk.channel_list}, neck: {self.neck.backbone_channel_list}"
+
+    def forward(self, sample: torch.Tensor):
+        # Forward through backbone
+        features, pos = self.neck(self.trunk(sample))
+        if self.scalp > 0:
+            # Discard the lowest resolution features
+            features, pos = features[: -self.scalp], pos[: -self.scalp]
+
+        src = features[-1]
+        output = {
+            "vision_features": src,
+            "vision_pos_enc": pos,
+            "backbone_fpn": features,
+        }
+        return output
+
+
+class FpnNeck(nn.Module):
+    """
+    A modified variant of Feature Pyramid Network (FPN) neck
+    (we remove output conv and also do bicubic interpolation similar to ViT
+    pos embed interpolation)
+    """
+
+    def __init__(
+        self,
+        position_encoding: nn.Module,
+        d_model: int,
+        backbone_channel_list: List[int],
+        kernel_size: int = 1,
+        stride: int = 1,
+        padding: int = 0,
+        fpn_interp_model: str = "bilinear",
+        fuse_type: str = "sum",
+        fpn_top_down_levels: Optional[List[int]] = None,
+    ):
+        """Initialize the neck
+        :param trunk: the backbone
+        :param position_encoding: the positional encoding to use
+        :param d_model: the dimension of the model
+        :param neck_norm: the normalization to use
+        """
+        super().__init__()
+        self.position_encoding = position_encoding
+        self.convs = nn.ModuleList()
+        self.backbone_channel_list = backbone_channel_list
+        for dim in backbone_channel_list:
+            current = nn.Sequential()
+            current.add_module(
+                "conv",
+                nn.Conv2d(
+                    in_channels=dim,
+                    out_channels=d_model,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                ),
+            )
+
+            self.convs.append(current)
+        self.fpn_interp_model = fpn_interp_model
+        assert fuse_type in ["sum", "avg"]
+        self.fuse_type = fuse_type
+
+        # levels to have top-down features in its outputs
+        # e.g. if fpn_top_down_levels is [2, 3], then only outputs of level 2 and 3
+        # have top-down propagation, while outputs of level 0 and level 1 have only
+        # lateral features from the same backbone level.
+        if fpn_top_down_levels is None:
+            # default is to have top-down features on all levels
+            fpn_top_down_levels = range(len(self.convs))
+        self.fpn_top_down_levels = list(fpn_top_down_levels)
+
+    def forward(self, xs: List[torch.Tensor]):
+
+        out = [None] * len(self.convs)
+        pos = [None] * len(self.convs)
+        assert len(xs) == len(self.convs)
+        # fpn forward pass
+        # see https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/fpn.py
+        prev_features = None
+        # forward in top-down order (from low to high resolution)
+        n = len(self.convs) - 1
+        for i in range(n, -1, -1):
+            x = xs[i]
+            lateral_features = self.convs[n - i](x)
+            if i in self.fpn_top_down_levels and prev_features is not None:
+                top_down_features = F.interpolate(
+                    prev_features.to(dtype=torch.float32),
+                    scale_factor=2.0,
+                    mode=self.fpn_interp_model,
+                    align_corners=(
+                        None if self.fpn_interp_model == "nearest" else False
+                    ),
+                    antialias=False,
+                )
+                prev_features = lateral_features + top_down_features
+                if self.fuse_type == "avg":
+                    prev_features /= 2
+            else:
+                prev_features = lateral_features
+            x_out = prev_features
+            out[i] = x_out
+            pos[i] = self.position_encoding(x_out).to(x_out.dtype)
+
+        return out, pos

sam2/modeling/backbones/utils.py

@@ -0,0 +1,95 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Some utilities for backbones, in particular for windowing"""
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def window_partition(x, window_size):
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(windows, window_size, pad_hw, hw):
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, ...] = (7, 7),
+        stride: Tuple[int, ...] = (4, 4),
+        padding: Tuple[int, ...] = (3, 3),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ):
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

sam2/modeling/memory_attention.py

@@ -0,0 +1,169 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional
+
+import torch
+from torch import nn, Tensor
+
+from sam2.modeling.sam.transformer import RoPEAttention
+
+from sam2.modeling.sam2_utils import get_activation_fn, get_clones
+
+
+class MemoryAttentionLayer(nn.Module):
+
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        self_attention: nn.Module,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+
+        # Where to add pos enc
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+    def _forward_sa(self, tgt, query_pos):
+        # Self-Attention
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(q, k, v=tgt2)
+        tgt = tgt + self.dropout1(tgt2)
+        return tgt
+
+    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
+        kwds = {}
+        if num_k_exclude_rope > 0:
+            assert isinstance(self.cross_attn_image, RoPEAttention)
+            kwds = {"num_k_exclude_rope": num_k_exclude_rope}
+
+        # Cross-Attention
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            v=memory,
+            **kwds,
+        )
+        tgt = tgt + self.dropout2(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        num_k_exclude_rope: int = 0,
+    ) -> torch.Tensor:
+
+        # Self-Attn, Cross-Attn
+        tgt = self._forward_sa(tgt, query_pos)
+        tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+        # MLP
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+
+class MemoryAttention(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        pos_enc_at_input: bool,
+        layer: nn.Module,
+        num_layers: int,
+        batch_first: bool = True,  # Do layers expect batch first input?
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = nn.LayerNorm(d_model)
+        self.pos_enc_at_input = pos_enc_at_input
+        self.batch_first = batch_first
+
+    def forward(
+        self,
+        curr: torch.Tensor,  # self-attention inputs
+        memory: torch.Tensor,  # cross-attention inputs
+        curr_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
+        memory_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
+        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
+    ):
+        if isinstance(curr, list):
+            assert isinstance(curr_pos, list)
+            assert len(curr) == len(curr_pos) == 1
+            curr, curr_pos = (
+                curr[0],
+                curr_pos[0],
+            )
+
+        assert (
+            curr.shape[1] == memory.shape[1]
+        ), "Batch size must be the same for curr and memory"
+
+        output = curr
+        if self.pos_enc_at_input and curr_pos is not None:
+            output = output + 0.1 * curr_pos
+
+        if self.batch_first:
+            # Convert to batch first
+            output = output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+            memory = memory.transpose(0, 1)
+            memory_pos = memory_pos.transpose(0, 1)
+
+        for layer in self.layers:
+            kwds = {}
+            if isinstance(layer.cross_attn_image, RoPEAttention):
+                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
+
+            output = layer(
+                tgt=output,
+                memory=memory,
+                pos=memory_pos,
+                query_pos=curr_pos,
+                **kwds,
+            )
+        normed_output = self.norm(output)
+
+        if self.batch_first:
+            # Convert back to seq first
+            normed_output = normed_output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+
+        return normed_output

sam2/modeling/memory_encoder.py

@@ -0,0 +1,181 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from sam2.modeling.sam2_utils import DropPath, get_clones, LayerNorm2d
+
+
+class MaskDownSampler(nn.Module):
+    """
+    Progressively downsample a mask by total_stride, each time by stride.
+    Note that LayerNorm is applied per *token*, like in ViT.
+
+    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
+    In the end, we linearly project to embed_dim channels.
+    """
+
+    def __init__(
+        self,
+        embed_dim=256,
+        kernel_size=4,
+        stride=4,
+        padding=0,
+        total_stride=16,
+        activation=nn.GELU,
+    ):
+        super().__init__()
+        num_layers = int(math.log2(total_stride) // math.log2(stride))
+        assert stride**num_layers == total_stride
+        self.encoder = nn.Sequential()
+        mask_in_chans, mask_out_chans = 1, 1
+        for _ in range(num_layers):
+            mask_out_chans = mask_in_chans * (stride**2)
+            self.encoder.append(
+                nn.Conv2d(
+                    mask_in_chans,
+                    mask_out_chans,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                )
+            )
+            self.encoder.append(LayerNorm2d(mask_out_chans))
+            self.encoder.append(activation())
+            mask_in_chans = mask_out_chans
+
+        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
+
+    def forward(self, x):
+        return self.encoder(x)
+
+
+# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
+class CXBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+    """
+
+    def __init__(
+        self,
+        dim,
+        kernel_size=7,
+        padding=3,
+        drop_path=0.0,
+        layer_scale_init_value=1e-6,
+        use_dwconv=True,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv2d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=padding,
+            groups=dim if use_dwconv else 1,
+        )  # depthwise conv
+        self.norm = LayerNorm2d(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(
+            dim, 4 * dim
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = self.norm(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class Fuser(nn.Module):
+    def __init__(self, layer, num_layers, dim=None, input_projection=False):
+        super().__init__()
+        self.proj = nn.Identity()
+        self.layers = get_clones(layer, num_layers)
+
+        if input_projection:
+            assert dim is not None
+            self.proj = nn.Conv2d(dim, dim, kernel_size=1)
+
+    def forward(self, x):
+        # normally x: (N, C, H, W)
+        x = self.proj(x)
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class MemoryEncoder(nn.Module):
+    def __init__(
+        self,
+        out_dim,
+        mask_downsampler,
+        fuser,
+        position_encoding,
+        in_dim=256,  # in_dim of pix_feats
+    ):
+        super().__init__()
+
+        self.mask_downsampler = mask_downsampler
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+        self.fuser = fuser
+        self.position_encoding = position_encoding
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(
+        self,
+        pix_feat: torch.Tensor,
+        masks: torch.Tensor,
+        skip_mask_sigmoid: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        ## Process masks
+        # sigmoid, so that less domain shift from gt masks which are bool
+        if not skip_mask_sigmoid:
+            masks = F.sigmoid(masks)
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return {"vision_features": x, "vision_pos_enc": [pos]}

sam2/modeling/position_encoding.py

@@ -0,0 +1,216 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Any, Optional, Tuple
+
+import numpy as np
+
+import torch
+from torch import nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats,
+        temperature: int = 10000,
+        normalize: bool = True,
+        scale: Optional[float] = None,
+    ):
+        super().__init__()
+        assert num_pos_feats % 2 == 0, "Expecting even model width"
+        self.num_pos_feats = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+        self.cache = {}
+
+    def _encode_xy(self, x, y):
+        # The positions are expected to be normalized
+        assert len(x) == len(y) and x.ndim == y.ndim == 1
+        x_embed = x * self.scale
+        y_embed = y * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, None] / dim_t
+        pos_y = y_embed[:, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        pos_y = torch.stack(
+            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        return pos_x, pos_y
+
+    @torch.no_grad()
+    def encode_boxes(self, x, y, w, h):
+        pos_x, pos_y = self._encode_xy(x, y)
+        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+        return pos
+
+    encode = encode_boxes  # Backwards compatibility
+
+    @torch.no_grad()
+    def encode_points(self, x, y, labels):
+        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
+        assert bx == by and nx == ny and bx == bl and nx == nl
+        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
+        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
+        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
+        return pos
+
+    @torch.no_grad()
+    def forward(self, x: torch.Tensor):
+        cache_key = (x.shape[-2], x.shape[-1])
+        if cache_key in self.cache:
+            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
+        y_embed = (
+            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
+            .view(1, -1, 1)
+            .repeat(x.shape[0], 1, x.shape[-1])
+        )
+        x_embed = (
+            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
+            .view(1, 1, -1)
+            .repeat(x.shape[0], x.shape[-2], 1)
+        )
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        self.cache[cache_key] = pos[0]
+        return pos
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C
+
+
+# Rotary Positional Encoding, adapted from:
+# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+# 2. https://github.com/naver-ai/rope-vit
+# 3. https://github.com/lucidrains/rotary-embedding-torch
+
+
+def init_t_xy(end_x: int, end_y: int):
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x, t_y
+
+
+def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+):
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)

sam2/modeling/sam/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam2/modeling/sam/mask_decoder.py

@@ -0,0 +1,295 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from sam2.modeling.sam2_utils import LayerNorm2d, MLP
+
+
+class MaskDecoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+        use_high_res_features: bool = False,
+        iou_prediction_use_sigmoid=False,
+        dynamic_multimask_via_stability=False,
+        dynamic_multimask_stability_delta=0.05,
+        dynamic_multimask_stability_thresh=0.98,
+        pred_obj_scores: bool = False,
+        pred_obj_scores_mlp: bool = False,
+        use_multimask_token_for_obj_ptr: bool = False,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.pred_obj_scores = pred_obj_scores
+        if self.pred_obj_scores:
+            self.obj_score_token = nn.Embedding(1, transformer_dim)
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
+            ),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(
+                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+            activation(),
+        )
+        self.use_high_res_features = use_high_res_features
+        if use_high_res_features:
+            self.conv_s0 = nn.Conv2d(
+                transformer_dim, transformer_dim // 8, kernel_size=1, stride=1
+            )
+            self.conv_s1 = nn.Conv2d(
+                transformer_dim, transformer_dim // 4, kernel_size=1, stride=1
+            )
+
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [
+                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
+                for i in range(self.num_mask_tokens)
+            ]
+        )
+
+        self.iou_prediction_head = MLP(
+            transformer_dim,
+            iou_head_hidden_dim,
+            self.num_mask_tokens,
+            iou_head_depth,
+            sigmoid_output=iou_prediction_use_sigmoid,
+        )
+        if self.pred_obj_scores:
+            self.pred_obj_score_head = nn.Linear(transformer_dim, 1)
+            if pred_obj_scores_mlp:
+                self.pred_obj_score_head = MLP(transformer_dim, transformer_dim, 1, 3)
+
+        # When outputting a single mask, optionally we can dynamically fall back to the best
+        # multimask output token if the single mask output token gives low stability scores.
+        self.dynamic_multimask_via_stability = dynamic_multimask_via_stability
+        self.dynamic_multimask_stability_delta = dynamic_multimask_stability_delta
+        self.dynamic_multimask_stability_thresh = dynamic_multimask_stability_thresh
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+          torch.Tensor: batched SAM token for mask output
+        """
+        masks, iou_pred, mask_tokens_out, object_score_logits = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+            repeat_image=repeat_image,
+            high_res_features=high_res_features,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            masks = masks[:, 1:, :, :]
+            iou_pred = iou_pred[:, 1:]
+        elif self.dynamic_multimask_via_stability and not self.training:
+            masks, iou_pred = self._dynamic_multimask_via_stability(masks, iou_pred)
+        else:
+            masks = masks[:, 0:1, :, :]
+            iou_pred = iou_pred[:, 0:1]
+
+        if multimask_output and self.use_multimask_token_for_obj_ptr:
+            sam_tokens_out = mask_tokens_out[:, 1:]  # [b, 3, c] shape
+        else:
+            # Take the mask output token. Here we *always* use the token for single mask output.
+            # At test time, even if we track after 1-click (and using multimask_output=True),
+            # we still take the single mask token here. The rationale is that we always track
+            # after multiple clicks during training, so the past tokens seen during training
+            # are always the single mask token (and we'll let it be the object-memory token).
+            sam_tokens_out = mask_tokens_out[:, 0:1]  # [b, 1, c] shape
+
+        # Prepare output
+        return masks, iou_pred, sam_tokens_out, object_score_logits
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        s = 0
+        if self.pred_obj_scores:
+            output_tokens = torch.cat(
+                [
+                    self.obj_score_token.weight,
+                    self.iou_token.weight,
+                    self.mask_tokens.weight,
+                ],
+                dim=0,
+            )
+            s = 1
+        else:
+            output_tokens = torch.cat(
+                [self.iou_token.weight, self.mask_tokens.weight], dim=0
+            )
+        output_tokens = output_tokens.unsqueeze(0).expand(
+            sparse_prompt_embeddings.size(0), -1, -1
+        )
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        if repeat_image:
+            src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        else:
+            assert image_embeddings.shape[0] == tokens.shape[0]
+            src = image_embeddings
+        src = src + dense_prompt_embeddings
+        assert (
+            image_pe.size(0) == 1
+        ), "image_pe should have size 1 in batch dim (from `get_dense_pe()`)"
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, s, :]
+        mask_tokens_out = hs[:, s + 1 : (s + 1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        if not self.use_high_res_features:
+            upscaled_embedding = self.output_upscaling(src)
+        else:
+            dc1, ln1, act1, dc2, act2 = self.output_upscaling
+            feat_s0, feat_s1 = high_res_features
+            upscaled_embedding = act1(ln1(dc1(src) + feat_s1))
+            upscaled_embedding = act2(dc2(upscaled_embedding) + feat_s0)
+
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            hyper_in_list.append(
+                self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
+            )
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+        if self.pred_obj_scores:
+            assert s == 1
+            object_score_logits = self.pred_obj_score_head(hs[:, 0, :])
+        else:
+            # Obj scores logits - default to 10.0, i.e. assuming the object is present, sigmoid(10)=1
+            object_score_logits = 10.0 * iou_pred.new_ones(iou_pred.shape[0], 1)
+
+        return masks, iou_pred, mask_tokens_out, object_score_logits
+
+    def _get_stability_scores(self, mask_logits):
+        """
+        Compute stability scores of the mask logits based on the IoU between upper and
+        lower thresholds, similar to https://github.com/fairinternal/onevision/pull/568.
+        """
+        mask_logits = mask_logits.flatten(-2)
+        stability_delta = self.dynamic_multimask_stability_delta
+        area_i = torch.sum(mask_logits > stability_delta, dim=-1).float()
+        area_u = torch.sum(mask_logits > -stability_delta, dim=-1).float()
+        stability_scores = torch.where(area_u > 0, area_i / area_u, 1.0)
+        return stability_scores
+
+    def _dynamic_multimask_via_stability(self, all_mask_logits, all_iou_scores):
+        """
+        When outputting a single mask, if the stability score from the current single-mask
+        output (based on output token 0) falls below a threshold, we instead select from
+        multi-mask outputs (based on output token 1~3) the mask with the highest predicted
+        IoU score. This is intended to ensure a valid mask for both clicking and tracking.
+        """
+        # The best mask from multimask output tokens (1~3)
+        multimask_logits = all_mask_logits[:, 1:, :, :]
+        multimask_iou_scores = all_iou_scores[:, 1:]
+        best_scores_inds = torch.argmax(multimask_iou_scores, dim=-1)
+        batch_inds = torch.arange(
+            multimask_iou_scores.size(0), device=all_iou_scores.device
+        )
+        best_multimask_logits = multimask_logits[batch_inds, best_scores_inds]
+        best_multimask_logits = best_multimask_logits.unsqueeze(1)
+        best_multimask_iou_scores = multimask_iou_scores[batch_inds, best_scores_inds]
+        best_multimask_iou_scores = best_multimask_iou_scores.unsqueeze(1)
+
+        # The mask from singlemask output token 0 and its stability score
+        singlemask_logits = all_mask_logits[:, 0:1, :, :]
+        singlemask_iou_scores = all_iou_scores[:, 0:1]
+        stability_scores = self._get_stability_scores(singlemask_logits)
+        is_stable = stability_scores >= self.dynamic_multimask_stability_thresh
+
+        # Dynamically fall back to best multimask output upon low stability scores.
+        mask_logits_out = torch.where(
+            is_stable[..., None, None].expand_as(singlemask_logits),
+            singlemask_logits,
+            best_multimask_logits,
+        )
+        iou_scores_out = torch.where(
+            is_stable.expand_as(singlemask_iou_scores),
+            singlemask_iou_scores,
+            best_multimask_iou_scores,
+        )
+        return mask_logits_out, iou_scores_out

sam2/modeling/sam/prompt_encoder.py

@@ -0,0 +1,182 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from sam2.modeling.position_encoding import PositionEmbeddingRandom
+
+from sam2.modeling.sam2_utils import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [
+            nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)
+        ]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (
+            4 * image_embedding_size[0],
+            4 * image_embedding_size[1],
+        )
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(
+            points, self.input_image_size
+        )
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        point_embedding[labels == 2] += self.point_embeddings[2].weight
+        point_embedding[labels == 3] += self.point_embeddings[3].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(
+            coords, self.input_image_size
+        )
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty(
+            (bs, 0, self.embed_dim), device=self._get_device()
+        )
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings

sam2/modeling/sam/transformer.py

@@ -0,0 +1,360 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import contextlib
+import math
+import warnings
+from functools import partial
+from typing import Tuple, Type
+
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+
+from sam2.modeling.position_encoding import apply_rotary_enc, compute_axial_cis
+from sam2.modeling.sam2_utils import MLP
+from sam2.utils.misc import get_sdpa_settings
+
+warnings.simplefilter(action="ignore", category=FutureWarning)
+# Check whether Flash Attention is available (and use it by default)
+OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
+# A fallback setting to allow all available kernels if Flash Attention fails
+ALLOW_ALL_KERNELS = False
+
+
+def sdp_kernel_context(dropout_p):
+    """
+    Get the context for the attention scaled dot-product kernel. We use Flash Attention
+    by default, but fall back to all available kernels if Flash Attention fails.
+    """
+    if ALLOW_ALL_KERNELS:
+        return contextlib.nullcontext()
+
+    return torch.backends.cuda.sdp_kernel(
+        enable_flash=USE_FLASH_ATTN,
+        # if Flash attention kernel is off, then math kernel needs to be enabled
+        enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+        enable_mem_efficient=OLD_GPU,
+    )
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attention layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLP(
+            embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=activation
+        )
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(
+        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+        dropout: float = 0.0,
+        kv_in_dim: int = None,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert (
+            self.internal_dim % num_heads == 0
+        ), "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+        self.dropout_p = dropout
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        try:
+            with sdp_kernel_context(dropout_p):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        except Exception as e:
+            # Fall back to all kernels if the Flash attention kernel fails
+            warnings.warn(
+                f"Flash Attention kernel failed due to: {e}\nFalling back to all available "
+                f"kernels for scaled_dot_product_attention (which may have a slower speed).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            global ALLOW_ALL_KERNELS
+            ALLOW_ALL_KERNELS = True
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out
+
+
+class RoPEAttention(Attention):
+    """Attention with rotary position encoding."""
+
+    def __init__(
+        self,
+        *args,
+        rope_theta=10000.0,
+        # whether to repeat q rope to match k length
+        # this is needed for cross-attention to memories
+        rope_k_repeat=False,
+        feat_sizes=(32, 32),  # [w, h] for stride 16 feats at 512 resolution
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.compute_cis = partial(
+            compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
+        )
+        freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
+        self.freqs_cis = freqs_cis
+        self.rope_k_repeat = rope_k_repeat
+
+    def forward(
+        self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
+    ) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Apply rotary position encoding
+        w = h = math.sqrt(q.shape[-2])
+        self.freqs_cis = self.freqs_cis.to(q.device)
+        if self.freqs_cis.shape[0] != q.shape[-2]:
+            self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
+        if q.shape[-2] != k.shape[-2]:
+            assert self.rope_k_repeat
+
+        num_k_rope = k.size(-2) - num_k_exclude_rope
+        q, k[:, :, :num_k_rope] = apply_rotary_enc(
+            q,
+            k[:, :, :num_k_rope],
+            freqs_cis=self.freqs_cis,
+            repeat_freqs_k=self.rope_k_repeat,
+        )
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        try:
+            with sdp_kernel_context(dropout_p):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        except Exception as e:
+            # Fall back to all kernels if the Flash attention kernel fails
+            warnings.warn(
+                f"Flash Attention kernel failed due to: {e}\nFalling back to all available "
+                f"kernels for scaled_dot_product_attention (which may have a slower speed).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            global ALLOW_ALL_KERNELS
+            ALLOW_ALL_KERNELS = True
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

sam2/modeling/sam2_base.py

@@ -0,0 +1,829 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.distributed
+import torch.nn.functional as F
+
+from torch.nn.init import trunc_normal_
+
+from sam2.modeling.sam.mask_decoder import MaskDecoder
+from sam2.modeling.sam.prompt_encoder import PromptEncoder
+from sam2.modeling.sam.transformer import TwoWayTransformer
+from sam2.modeling.sam2_utils import get_1d_sine_pe, MLP, select_closest_cond_frames
+
+# a large negative value as a placeholder score for missing objects
+NO_OBJ_SCORE = -1024.0
+
+
+class SAM2Base(torch.nn.Module):
+    def __init__(
+        self,
+        image_encoder,
+        memory_attention,
+        memory_encoder,
+        num_maskmem=7,  # default 1 input frame + 6 previous frames
+        image_size=512,
+        backbone_stride=16,  # stride of the image backbone output
+        sigmoid_scale_for_mem_enc=1.0,  # scale factor for mask sigmoid prob
+        sigmoid_bias_for_mem_enc=0.0,  # bias factor for mask sigmoid prob
+        # During evaluation, whether to binarize the sigmoid mask logits on interacted frames with clicks
+        binarize_mask_from_pts_for_mem_enc=False,
+        use_mask_input_as_output_without_sam=False,  # on frames with mask input, whether to directly output the input mask without using a SAM prompt encoder + mask decoder
+        # The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
+        # we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
+        # a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
+        max_cond_frames_in_attn=-1,
+        # on the first frame, whether to directly add the no-memory embedding to the image feature
+        # (instead of using the transformer encoder)
+        directly_add_no_mem_embed=False,
+        # whether to use high-resolution feature maps in the SAM mask decoder
+        use_high_res_features_in_sam=False,
+        # whether to output multiple (3) masks for the first click on initial conditioning frames
+        multimask_output_in_sam=False,
+        # the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
+        # default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
+        multimask_min_pt_num=1,
+        multimask_max_pt_num=1,
+        # whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
+        multimask_output_for_tracking=False,
+        # Whether to use multimask tokens for obj ptr; Only relevant when both
+        # use_obj_ptrs_in_encoder=True and multimask_output_for_tracking=True
+        use_multimask_token_for_obj_ptr: bool = False,
+        # whether to use sigmoid to restrict ious prediction to [0-1]
+        iou_prediction_use_sigmoid=False,
+        # The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
+        # For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
+        # (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
+        memory_temporal_stride_for_eval=1,
+        # if `add_all_frames_to_correct_as_cond` is True, we also append to the conditioning frame list any frame that receives a later correction click
+        # if `add_all_frames_to_correct_as_cond` is False, we conditioning frame list to only use those initial conditioning frames
+        add_all_frames_to_correct_as_cond=False,
+        # whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
+        non_overlap_masks_for_mem_enc=False,
+        # whether to cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+        use_obj_ptrs_in_encoder=False,
+        # the maximum number of object pointers from other frames in encoder cross attention (only relevant when `use_obj_ptrs_in_encoder=True`)
+        max_obj_ptrs_in_encoder=16,
+        # whether to add temporal positional encoding to the object pointers in the encoder (only relevant when `use_obj_ptrs_in_encoder=True`)
+        add_tpos_enc_to_obj_ptrs=True,
+        # whether to add an extra linear projection layer for the temporal positional encoding in the object pointers to avoid potential interference
+        # with spatial positional encoding (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
+        proj_tpos_enc_in_obj_ptrs=False,
+        # whether to only attend to object pointers in the past (before the current frame) in the encoder during evaluation
+        # (only relevant when `use_obj_ptrs_in_encoder=True`; this might avoid pointer information too far in the future to distract the initial tracking)
+        only_obj_ptrs_in_the_past_for_eval=False,
+        # Whether to predict if there is an object in the frame
+        pred_obj_scores: bool = False,
+        # Whether to use an MLP to predict object scores
+        pred_obj_scores_mlp: bool = False,
+        # Only relevant if pred_obj_scores=True and use_obj_ptrs_in_encoder=True;
+        # Whether to have a fixed no obj pointer when there is no object present
+        # or to use it as an additive embedding with obj_ptr produced by decoder
+        fixed_no_obj_ptr: bool = False,
+        # Soft no object, i.e. mix in no_obj_ptr softly,
+        # hope to make recovery easier if there is a mistake and mitigate accumulation of errors
+        soft_no_obj_ptr: bool = False,
+        use_mlp_for_obj_ptr_proj: bool = False,
+        # extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
+        sam_mask_decoder_extra_args=None,
+        compile_image_encoder: bool = False,
+    ):
+        super().__init__()
+
+        # Part 1: the image backbone
+        self.image_encoder = image_encoder
+        # Use level 0, 1, 2 for high-res setting, or just level 2 for the default setting
+        self.use_high_res_features_in_sam = use_high_res_features_in_sam
+        self.num_feature_levels = 3 if use_high_res_features_in_sam else 1
+        self.use_obj_ptrs_in_encoder = use_obj_ptrs_in_encoder
+        self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
+        if use_obj_ptrs_in_encoder:
+            # A conv layer to downsample the mask prompt to stride 4 (the same stride as
+            # low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
+            # so that it can be fed into the SAM mask decoder to generate a pointer.
+            self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
+        self.add_tpos_enc_to_obj_ptrs = add_tpos_enc_to_obj_ptrs
+        if proj_tpos_enc_in_obj_ptrs:
+            assert add_tpos_enc_to_obj_ptrs  # these options need to be used together
+        self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
+        self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval
+
+        # Part 2: memory attention to condition current frame's visual features
+        # with memories (and obj ptrs) from past frames
+        self.memory_attention = memory_attention
+        self.hidden_dim = memory_attention.d_model
+
+        # Part 3: memory encoder for the previous frame's outputs
+        self.memory_encoder = memory_encoder
+        self.mem_dim = self.hidden_dim
+        if hasattr(self.memory_encoder, "out_proj") and hasattr(
+            self.memory_encoder.out_proj, "weight"
+        ):
+            # if there is compression of memories along channel dim
+            self.mem_dim = self.memory_encoder.out_proj.weight.shape[0]
+        self.num_maskmem = num_maskmem  # Number of memories accessible
+        # Temporal encoding of the memories
+        self.maskmem_tpos_enc = torch.nn.Parameter(
+            torch.zeros(num_maskmem, 1, 1, self.mem_dim)
+        )
+        trunc_normal_(self.maskmem_tpos_enc, std=0.02)
+        # a single token to indicate no memory embedding from previous frames
+        self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        trunc_normal_(self.no_mem_embed, std=0.02)
+        trunc_normal_(self.no_mem_pos_enc, std=0.02)
+        self.directly_add_no_mem_embed = directly_add_no_mem_embed
+        # Apply sigmoid to the output raw mask logits (to turn them from
+        # range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
+        self.sigmoid_scale_for_mem_enc = sigmoid_scale_for_mem_enc
+        self.sigmoid_bias_for_mem_enc = sigmoid_bias_for_mem_enc
+        self.binarize_mask_from_pts_for_mem_enc = binarize_mask_from_pts_for_mem_enc
+        self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
+        self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
+        # On frames with mask input, whether to directly output the input mask without
+        # using a SAM prompt encoder + mask decoder
+        self.use_mask_input_as_output_without_sam = use_mask_input_as_output_without_sam
+        self.multimask_output_in_sam = multimask_output_in_sam
+        self.multimask_min_pt_num = multimask_min_pt_num
+        self.multimask_max_pt_num = multimask_max_pt_num
+        self.multimask_output_for_tracking = multimask_output_for_tracking
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+        self.iou_prediction_use_sigmoid = iou_prediction_use_sigmoid
+
+        # Part 4: SAM-style prompt encoder (for both mask and point inputs)
+        # and SAM-style mask decoder for the final mask output
+        self.image_size = image_size
+        self.backbone_stride = backbone_stride
+        self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
+        self.pred_obj_scores = pred_obj_scores
+        self.pred_obj_scores_mlp = pred_obj_scores_mlp
+        self.fixed_no_obj_ptr = fixed_no_obj_ptr
+        self.soft_no_obj_ptr = soft_no_obj_ptr
+        if self.fixed_no_obj_ptr:
+            assert self.pred_obj_scores
+            assert self.use_obj_ptrs_in_encoder
+        if self.pred_obj_scores and self.use_obj_ptrs_in_encoder:
+            self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
+            trunc_normal_(self.no_obj_ptr, std=0.02)
+        self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj
+
+        self._build_sam_heads()
+        self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
+        self.max_cond_frames_in_attn = max_cond_frames_in_attn
+
+        # Model compilation
+        if compile_image_encoder:
+            # Compile the forward function (not the full module) to allow loading checkpoints.
+            print(
+                "Image encoder compilation is enabled. First forward pass will be slow."
+            )
+            self.image_encoder.forward = torch.compile(
+                self.image_encoder.forward,
+                mode="max-autotune",
+                fullgraph=True,
+                dynamic=False,
+            )
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError(
+            "Please use the corresponding methods in SAM2VideoPredictor for inference."
+            "See notebooks/video_predictor_example.ipynb for an example."
+        )
+
+    def _build_sam_heads(self):
+        """Build SAM-style prompt encoder and mask decoder."""
+        self.sam_prompt_embed_dim = self.hidden_dim
+        self.sam_image_embedding_size = self.image_size // self.backbone_stride
+
+        # build PromptEncoder and MaskDecoder from SAM
+        # (their hyperparameters like `mask_in_chans=16` are from SAM code)
+        self.sam_prompt_encoder = PromptEncoder(
+            embed_dim=self.sam_prompt_embed_dim,
+            image_embedding_size=(
+                self.sam_image_embedding_size,
+                self.sam_image_embedding_size,
+            ),
+            input_image_size=(self.image_size, self.image_size),
+            mask_in_chans=16,
+        )
+        self.sam_mask_decoder = MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=self.sam_prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=self.sam_prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            use_high_res_features=self.use_high_res_features_in_sam,
+            iou_prediction_use_sigmoid=self.iou_prediction_use_sigmoid,
+            pred_obj_scores=self.pred_obj_scores,
+            pred_obj_scores_mlp=self.pred_obj_scores_mlp,
+            use_multimask_token_for_obj_ptr=self.use_multimask_token_for_obj_ptr,
+            **(self.sam_mask_decoder_extra_args or {}),
+        )
+        if self.use_obj_ptrs_in_encoder:
+            # a linear projection on SAM output tokens to turn them into object pointers
+            self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
+            if self.use_mlp_for_obj_ptr_proj:
+                self.obj_ptr_proj = MLP(
+                    self.hidden_dim, self.hidden_dim, self.hidden_dim, 3
+                )
+        else:
+            self.obj_ptr_proj = torch.nn.Identity()
+        if self.proj_tpos_enc_in_obj_ptrs:
+            # a linear projection on temporal positional encoding in object pointers to
+            # avoid potential interference with spatial positional encoding
+            self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
+        else:
+            self.obj_ptr_tpos_proj = torch.nn.Identity()
+
+    def _forward_sam_heads(
+        self,
+        backbone_features,
+        point_inputs=None,
+        mask_inputs=None,
+        high_res_features=None,
+        multimask_output=False,
+    ):
+        """
+        Forward SAM prompt encoders and mask heads.
+
+        Inputs:
+        - backbone_features: image features of [B, C, H, W] shape
+        - point_inputs: a dictionary with "point_coords" and "point_labels", where
+          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
+             absolute pixel-unit coordinate in (x, y) format of the P input points
+          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
+             positive clicks, 0 means negative clicks, and -1 means padding
+        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
+          same spatial size as the image.
+        - high_res_features: either 1) None or 2) or a list of length 2 containing
+          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
+          which will be used as high-resolution feature maps for SAM decoder.
+        - multimask_output: if it's True, we output 3 candidate masks and their 3
+          corresponding IoU estimates, and if it's False, we output only 1 mask and
+          its corresponding IoU estimate.
+
+        Outputs:
+        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
+          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
+          output mask logits (before sigmoid) for the low-resolution masks, with 4x
+          the resolution (1/4 stride) of the input backbone_features.
+        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
+          if `multimask_output=True` and M = 1 if `multimask_output=False`),
+          upsampled from the low-resolution masks, with shape size as the image
+          (stride is 1 pixel).
+        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
+          if `multimask_output=False`), the estimated IoU of each output mask.
+        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `low_res_multimasks`.
+        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `high_res_multimasks`.
+        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
+          based on the output token from the SAM mask decoder.
+        """
+        B = backbone_features.size(0)
+        device = backbone_features.device
+        assert backbone_features.size(1) == self.sam_prompt_embed_dim
+        assert backbone_features.size(2) == self.sam_image_embedding_size
+        assert backbone_features.size(3) == self.sam_image_embedding_size
+
+        # a) Handle point prompts
+        if point_inputs is not None:
+            sam_point_coords = point_inputs["point_coords"]
+            sam_point_labels = point_inputs["point_labels"]
+            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
+        else:
+            # If no points are provide, pad with an empty point (with label -1)
+            sam_point_coords = torch.zeros(B, 1, 2, device=device)
+            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
+
+        # b) Handle mask prompts
+        if mask_inputs is not None:
+            # If mask_inputs is provided, downsize it into low-res mask input if needed
+            # and feed it as a dense mask prompt into the SAM mask encoder
+            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
+            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
+                sam_mask_prompt = F.interpolate(
+                    mask_inputs.float(),
+                    size=self.sam_prompt_encoder.mask_input_size,
+                    align_corners=False,
+                    mode="bilinear",
+                    antialias=True,  # use antialias for downsampling
+                )
+            else:
+                sam_mask_prompt = mask_inputs
+        else:
+            # Otherwise, simply feed None (and SAM's prompt encoder will add
+            # a learned `no_mask_embed` to indicate no mask input in this case).
+            sam_mask_prompt = None
+
+        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
+            points=(sam_point_coords, sam_point_labels),
+            boxes=None,
+            masks=sam_mask_prompt,
+        )
+        (
+            low_res_multimasks,
+            ious,
+            sam_output_tokens,
+            object_score_logits,
+        ) = self.sam_mask_decoder(
+            image_embeddings=backbone_features,
+            image_pe=self.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=False,  # the image is already batched
+            high_res_features=high_res_features,
+        )
+        if self.pred_obj_scores:
+            is_obj_appearing = object_score_logits > 0
+
+            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
+            # consistent with the actual mask prediction
+            low_res_multimasks = torch.where(
+                is_obj_appearing[:, None, None],
+                low_res_multimasks,
+                NO_OBJ_SCORE,
+            )
+
+        # convert masks from possibly bfloat16 (or float16) to float32
+        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
+        low_res_multimasks = low_res_multimasks.float()
+        high_res_multimasks = F.interpolate(
+            low_res_multimasks,
+            size=(self.image_size, self.image_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        sam_output_token = sam_output_tokens[:, 0]
+        if multimask_output:
+            # take the best mask prediction (with the highest IoU estimation)
+            best_iou_inds = torch.argmax(ious, dim=-1)
+            batch_inds = torch.arange(B, device=device)
+            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            if sam_output_tokens.size(1) > 1:
+                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
+        else:
+            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
+
+        # Extract object pointer from the SAM output token (with occlusion handling)
+        obj_ptr = self.obj_ptr_proj(sam_output_token)
+        if self.pred_obj_scores:
+            # Allow *soft* no obj ptr, unlike for masks
+            if self.soft_no_obj_ptr:
+                # Only hard possible with gt
+                assert not self.teacher_force_obj_scores_for_mem
+                lambda_is_obj_appearing = object_score_logits.sigmoid()
+            else:
+                lambda_is_obj_appearing = is_obj_appearing.float()
+
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_multimasks,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
+        """
+        Directly turn binary `mask_inputs` into a output mask logits without using SAM.
+        (same input and output shapes as in _forward_sam_heads above).
+        """
+        # Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
+        out_scale, out_bias = 20.0, -10.0  # sigmoid(-10.0)=4.5398e-05
+        mask_inputs_float = mask_inputs.float()
+        high_res_masks = mask_inputs_float * out_scale + out_bias
+        low_res_masks = F.interpolate(
+            high_res_masks,
+            size=(high_res_masks.size(-2) // 4, high_res_masks.size(-1) // 4),
+            align_corners=False,
+            mode="bilinear",
+            antialias=True,  # use antialias for downsampling
+        )
+        # a dummy IoU prediction of all 1's under mask input
+        ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
+        if not self.use_obj_ptrs_in_encoder:
+            # all zeros as a dummy object pointer (of shape [B, C])
+            obj_ptr = torch.zeros(
+                mask_inputs.size(0), self.hidden_dim, device=mask_inputs.device
+            )
+        else:
+            # produce an object pointer using the SAM decoder from the mask input
+            _, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
+                backbone_features=backbone_features,
+                mask_inputs=self.mask_downsample(mask_inputs_float),
+                high_res_features=high_res_features,
+            )
+        # In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
+        # Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
+        # on the object_scores from the SAM decoder.
+        is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
+        is_obj_appearing = is_obj_appearing[..., None]
+        lambda_is_obj_appearing = is_obj_appearing.float()
+        object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
+        if self.pred_obj_scores:
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_masks,
+            high_res_masks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def forward_image(self, img_batch: torch.Tensor):
+        """Get the image feature on the input batch."""
+        backbone_out = self.image_encoder(img_batch)
+        if self.use_high_res_features_in_sam:
+            # precompute projected level 0 and level 1 features in SAM decoder
+            # to avoid running it again on every SAM click
+            backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
+                backbone_out["backbone_fpn"][0]
+            )
+            backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
+                backbone_out["backbone_fpn"][1]
+            )
+        return backbone_out
+
+    def _prepare_backbone_features(self, backbone_out):
+        """Prepare and flatten visual features."""
+        backbone_out = backbone_out.copy()
+        assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
+        assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels
+
+        feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+        vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+
+        feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
+        # flatten NxCxHxW to HWxNxC
+        vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
+        vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]
+
+        return backbone_out, vision_feats, vision_pos_embeds, feat_sizes
+
+    def _prepare_memory_conditioned_features(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+    ):
+        """Fuse the current frame's visual feature map with previous memory."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        device = current_vision_feats[-1].device
+        # The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
+        # In this case, we skip the fusion with any memory.
+        if self.num_maskmem == 0:  # Disable memory and skip fusion
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+            return pix_feat
+
+        num_obj_ptr_tokens = 0
+        # Step 1: condition the visual features of the current frame on previous memories
+        if not is_init_cond_frame:
+            # Retrieve the memories encoded with the maskmem backbone
+            to_cat_memory, to_cat_memory_pos_embed = [], []
+            # Add conditioning frames's output first (all cond frames have t_pos=0 for
+            # when getting temporal positional embedding below)
+            assert len(output_dict["cond_frame_outputs"]) > 0
+            # Select a maximum number of temporally closest cond frames for cross attention
+            cond_outputs = output_dict["cond_frame_outputs"]
+            selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
+                frame_idx, cond_outputs, self.max_cond_frames_in_attn
+            )
+            t_pos_and_prevs = [(0, out) for out in selected_cond_outputs.values()]
+            # Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
+            # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
+            # We also allow taking the memory frame non-consecutively (with r>1), in which case
+            # we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
+            r = self.memory_temporal_stride_for_eval
+            for t_pos in range(1, self.num_maskmem):
+                t_rel = self.num_maskmem - t_pos  # how many frames before current frame
+                if t_rel == 1:
+                    # for t_rel == 1, we take the last frame (regardless of r)
+                    if not track_in_reverse:
+                        # the frame immediately before this frame (i.e. frame_idx - 1)
+                        prev_frame_idx = frame_idx - t_rel
+                    else:
+                        # the frame immediately after this frame (i.e. frame_idx + 1)
+                        prev_frame_idx = frame_idx + t_rel
+                else:
+                    # for t_rel >= 2, we take the memory frame from every r-th frames
+                    if not track_in_reverse:
+                        # first find the nearest frame among every r-th frames before this frame
+                        # for r=1, this would be (frame_idx - 2)
+                        prev_frame_idx = ((frame_idx - 2) // r) * r
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx - (t_rel - 2) * r
+                    else:
+                        # first find the nearest frame among every r-th frames after this frame
+                        # for r=1, this would be (frame_idx + 2)
+                        prev_frame_idx = -(-(frame_idx + 2) // r) * r
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx + (t_rel - 2) * r
+                out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
+                if out is None:
+                    # If an unselected conditioning frame is among the last (self.num_maskmem - 1)
+                    # frames, we still attend to it as if it's a non-conditioning frame.
+                    out = unselected_cond_outputs.get(prev_frame_idx, None)
+                t_pos_and_prevs.append((t_pos, out))
+
+            for t_pos, prev in t_pos_and_prevs:
+                if prev is None:
+                    continue  # skip padding frames
+                # "maskmem_features" might have been offloaded to CPU in demo use cases,
+                # so we load it back to GPU (it's a no-op if it's already on GPU).
+                feats = prev["maskmem_features"].cuda(non_blocking=True)
+                to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
+                # Spatial positional encoding (it might have been offloaded to CPU in eval)
+                maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
+                maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
+                # Temporal positional encoding
+                maskmem_enc = (
+                    maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t_pos - 1]
+                )
+                to_cat_memory_pos_embed.append(maskmem_enc)
+
+            # Construct the list of past object pointers
+            if self.use_obj_ptrs_in_encoder:
+                max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
+                # First add those object pointers from selected conditioning frames
+                # (optionally, only include object pointers in the past during evaluation)
+                if not self.training and self.only_obj_ptrs_in_the_past_for_eval:
+                    ptr_cond_outputs = {
+                        t: out
+                        for t, out in selected_cond_outputs.items()
+                        if (t >= frame_idx if track_in_reverse else t <= frame_idx)
+                    }
+                else:
+                    ptr_cond_outputs = selected_cond_outputs
+                pos_and_ptrs = [
+                    # Temporal pos encoding contains how far away each pointer is from current frame
+                    (abs(frame_idx - t), out["obj_ptr"])
+                    for t, out in ptr_cond_outputs.items()
+                ]
+                # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
+                for t_diff in range(1, max_obj_ptrs_in_encoder):
+                    t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
+                    if t < 0 or (num_frames is not None and t >= num_frames):
+                        break
+                    out = output_dict["non_cond_frame_outputs"].get(
+                        t, unselected_cond_outputs.get(t, None)
+                    )
+                    if out is not None:
+                        pos_and_ptrs.append((t_diff, out["obj_ptr"]))
+                # If we have at least one object pointer, add them to the across attention
+                if len(pos_and_ptrs) > 0:
+                    pos_list, ptrs_list = zip(*pos_and_ptrs)
+                    # stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
+                    obj_ptrs = torch.stack(ptrs_list, dim=0)
+                    # a temporal positional embedding based on how far each object pointer is from
+                    # the current frame (sine embedding normalized by the max pointer num).
+                    if self.add_tpos_enc_to_obj_ptrs:
+                        t_diff_max = max_obj_ptrs_in_encoder - 1
+                        tpos_dim = C if self.proj_tpos_enc_in_obj_ptrs else self.mem_dim
+                        obj_pos = torch.tensor(pos_list, device=device)
+                        obj_pos = get_1d_sine_pe(obj_pos / t_diff_max, dim=tpos_dim)
+                        obj_pos = self.obj_ptr_tpos_proj(obj_pos)
+                        obj_pos = obj_pos.unsqueeze(1).expand(-1, B, self.mem_dim)
+                    else:
+                        obj_pos = obj_ptrs.new_zeros(len(pos_list), B, self.mem_dim)
+                    if self.mem_dim < C:
+                        # split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
+                        obj_ptrs = obj_ptrs.reshape(
+                            -1, B, C // self.mem_dim, self.mem_dim
+                        )
+                        obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
+                        obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
+                    to_cat_memory.append(obj_ptrs)
+                    to_cat_memory_pos_embed.append(obj_pos)
+                    num_obj_ptr_tokens = obj_ptrs.shape[0]
+                else:
+                    num_obj_ptr_tokens = 0
+        else:
+            # for initial conditioning frames, encode them without using any previous memory
+            if self.directly_add_no_mem_embed:
+                # directly add no-mem embedding (instead of using the transformer encoder)
+                pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
+                pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+                return pix_feat_with_mem
+
+            # Use a dummy token on the first frame (to avoid emtpy memory input to tranformer encoder)
+            to_cat_memory = [self.no_mem_embed.expand(1, B, self.mem_dim)]
+            to_cat_memory_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]
+
+        # Step 2: Concatenate the memories and forward through the transformer encoder
+        memory = torch.cat(to_cat_memory, dim=0)
+        memory_pos_embed = torch.cat(to_cat_memory_pos_embed, dim=0)
+
+        pix_feat_with_mem = self.memory_attention(
+            curr=current_vision_feats,
+            curr_pos=current_vision_pos_embeds,
+            memory=memory,
+            memory_pos=memory_pos_embed,
+            num_obj_ptr_tokens=num_obj_ptr_tokens,
+        )
+        # reshape the output (HW)BC => BCHW
+        pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+        return pix_feat_with_mem
+
+    def _encode_new_memory(
+        self,
+        current_vision_feats,
+        feat_sizes,
+        pred_masks_high_res,
+        is_mask_from_pts,
+    ):
+        """Encode the current image and its prediction into a memory feature."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        # top-level feature, (HW)BC => BCHW
+        pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+        if self.non_overlap_masks_for_mem_enc and not self.training:
+            # optionally, apply non-overlapping constraints to the masks (it's applied
+            # in the batch dimension and should only be used during eval, where all
+            # the objects come from the same video under batch size 1).
+            pred_masks_high_res = self._apply_non_overlapping_constraints(
+                pred_masks_high_res
+            )
+        # scale the raw mask logits with a temperature before applying sigmoid
+        binarize = self.binarize_mask_from_pts_for_mem_enc and is_mask_from_pts
+        if binarize and not self.training:
+            mask_for_mem = (pred_masks_high_res > 0).float()
+        else:
+            # apply sigmoid on the raw mask logits to turn them into range (0, 1)
+            mask_for_mem = torch.sigmoid(pred_masks_high_res)
+        # apply scale and bias terms to the sigmoid probabilities
+        if self.sigmoid_scale_for_mem_enc != 1.0:
+            mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
+        if self.sigmoid_bias_for_mem_enc != 0.0:
+            mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
+        maskmem_out = self.memory_encoder(
+            pix_feat, mask_for_mem, skip_mask_sigmoid=True  # sigmoid already applied
+        )
+        maskmem_features = maskmem_out["vision_features"]
+        maskmem_pos_enc = maskmem_out["vision_pos_enc"]
+
+        return maskmem_features, maskmem_pos_enc
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+    ):
+        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
+        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
+        if len(current_vision_feats) > 1:
+            high_res_features = [
+                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
+            ]
+        else:
+            high_res_features = None
+        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
+            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
+            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
+            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
+            sam_outputs = self._use_mask_as_output(
+                pix_feat, high_res_features, mask_inputs
+            )
+        else:
+            # fused the visual feature with previous memory features in the memory bank
+            pix_feat_with_mem = self._prepare_memory_conditioned_features(
+                frame_idx=frame_idx,
+                is_init_cond_frame=is_init_cond_frame,
+                current_vision_feats=current_vision_feats[-1:],
+                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
+                feat_sizes=feat_sizes[-1:],
+                output_dict=output_dict,
+                num_frames=num_frames,
+                track_in_reverse=track_in_reverse,
+            )
+            # apply SAM-style segmentation head
+            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
+            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
+            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
+            if prev_sam_mask_logits is not None:
+                assert point_inputs is not None and mask_inputs is None
+                mask_inputs = prev_sam_mask_logits
+            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
+            sam_outputs = self._forward_sam_heads(
+                backbone_features=pix_feat_with_mem,
+                point_inputs=point_inputs,
+                mask_inputs=mask_inputs,
+                high_res_features=high_res_features,
+                multimask_output=multimask_output,
+            )
+        (
+            _,
+            _,
+            _,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            _,
+        ) = sam_outputs
+
+        current_out["pred_masks"] = low_res_masks
+        current_out["pred_masks_high_res"] = high_res_masks
+        current_out["obj_ptr"] = obj_ptr
+
+        # Finally run the memory encoder on the predicted mask to encode
+        # it into a new memory feature (that can be used in future frames)
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                is_mask_from_pts=(point_inputs is not None),
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+        return current_out
+
+    def _use_multimask(self, is_init_cond_frame, point_inputs):
+        """Whether to use multimask output in the SAM head."""
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        multimask_output = (
+            self.multimask_output_in_sam
+            and (is_init_cond_frame or self.multimask_output_for_tracking)
+            and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
+        )
+        return multimask_output
+
+    def _apply_non_overlapping_constraints(self, pred_masks):
+        """
+        Apply non-overlapping constraints to the object scores in pred_masks. Here we
+        keep only the highest scoring object at each spatial location in pred_masks.
+        """
+        batch_size = pred_masks.size(0)
+        if batch_size == 1:
+            return pred_masks
+
+        device = pred_masks.device
+        # "max_obj_inds": object index of the object with the highest score at each location
+        max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
+        # "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
+        batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
+        keep = max_obj_inds == batch_obj_inds
+        # suppress overlapping regions' scores below -10.0 so that the foreground regions
+        # don't overlap (here sigmoid(-10.0)=4.5398e-05)
+        pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
+        return pred_masks

sam2/modeling/sam2_utils.py

@@ -0,0 +1,149 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import copy
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+class DropPath(nn.Module):
+    # adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
+    def __init__(self, drop_prob=0.0, scale_by_keep=True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        if self.drop_prob == 0.0 or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        activation: nn.Module = nn.ReLU,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.sigmoid_output = sigmoid_output
+        self.act = activation()
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x

sam2/sam2_image_predictor.py

@@ -0,0 +1,463 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from PIL.Image import Image
+
+from sam2.modeling.sam2_base import SAM2Base
+
+from sam2.utils.transforms import SAM2Transforms
+
+
+class SAM2ImagePredictor:
+    def __init__(
+        self,
+        sam_model: SAM2Base,
+        mask_threshold=0.0,
+        max_hole_area=0.0,
+        max_sprinkle_area=0.0,
+    ) -> None:
+        """
+        Uses SAM-2 to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam-2): The model to use for mask prediction.
+          mask_threshold (float): The threshold to use when converting mask logits
+            to binary masks. Masks are thresholded at 0 by default.
+          fill_hole_area (int): If fill_hole_area > 0, we fill small holes in up to
+            the maximum area of fill_hole_area in low_res_masks.
+        """
+        super().__init__()
+        self.model = sam_model
+        self._transforms = SAM2Transforms(
+            resolution=self.model.image_size,
+            mask_threshold=mask_threshold,
+            max_hole_area=max_hole_area,
+            max_sprinkle_area=max_sprinkle_area,
+        )
+
+        # Predictor state
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        # Whether the predictor is set for single image or a batch of images
+        self._is_batch = False
+
+        # Predictor config
+        self.mask_threshold = mask_threshold
+
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (256, 256),
+            (128, 128),
+            (64, 64),
+        ]
+
+    @classmethod
+    def from_pretrained(cls, model_id: str, **kwargs) -> "SAM2ImagePredictor":
+        """
+        Load a pretrained model from the Hugging Face hub.
+
+        Arguments:
+          model_id (str): The Hugging Face repository ID.
+          **kwargs: Additional arguments to pass to the model constructor.
+
+        Returns:
+          (SAM2ImagePredictor): The loaded model.
+        """
+        from sam2.build_sam import build_sam2_hf
+
+        sam_model = build_sam2_hf(model_id, **kwargs)
+        return cls(sam_model)
+
+    @torch.no_grad()
+    def set_image(
+        self,
+        image: Union[np.ndarray, Image],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
+          with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        self.reset_predictor()
+        # Transform the image to the form expected by the model
+        if isinstance(image, np.ndarray):
+            logging.info("For numpy array image, we assume (HxWxC) format")
+            self._orig_hw = [image.shape[:2]]
+        elif isinstance(image, Image):
+            w, h = image.size
+            self._orig_hw = [(h, w)]
+        else:
+            raise NotImplementedError("Image format not supported")
+
+        input_image = self._transforms(image)
+        input_image = input_image[None, ...].to(self.device)
+
+        assert (
+            len(input_image.shape) == 4 and input_image.shape[1] == 3
+        ), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
+        logging.info("Computing image embeddings for the provided image...")
+        backbone_out = self.model.forward_image(input_image)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        logging.info("Image embeddings computed.")
+
+    @torch.no_grad()
+    def set_image_batch(
+        self,
+        image_list: List[Union[np.ndarray]],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image batch, allowing
+        masks to be predicted with the 'predict_batch' method.
+
+        Arguments:
+          image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
+          with pixel values in [0, 255].
+        """
+        self.reset_predictor()
+        assert isinstance(image_list, list)
+        self._orig_hw = []
+        for image in image_list:
+            assert isinstance(
+                image, np.ndarray
+            ), "Images are expected to be an np.ndarray in RGB format, and of shape  HWC"
+            self._orig_hw.append(image.shape[:2])
+        # Transform the image to the form expected by the model
+        img_batch = self._transforms.forward_batch(image_list)
+        img_batch = img_batch.to(self.device)
+        batch_size = img_batch.shape[0]
+        assert (
+            len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+        ), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+        logging.info("Computing image embeddings for the provided images...")
+        backbone_out = self.model.forward_image(img_batch)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        self._is_batch = True
+        logging.info("Image embeddings computed.")
+
+    def predict_batch(
+        self,
+        point_coords_batch: List[np.ndarray] = None,
+        point_labels_batch: List[np.ndarray] = None,
+        box_batch: List[np.ndarray] = None,
+        mask_input_batch: List[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        """This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
+        It returns a tupele of lists of masks, ious, and low_res_masks_logits.
+        """
+        assert self._is_batch, "This function should only be used when in batched mode"
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image_batch(...) before mask prediction."
+            )
+        num_images = len(self._features["image_embed"])
+        all_masks = []
+        all_ious = []
+        all_low_res_masks = []
+        for img_idx in range(num_images):
+            # Transform input prompts
+            point_coords = (
+                point_coords_batch[img_idx] if point_coords_batch is not None else None
+            )
+            point_labels = (
+                point_labels_batch[img_idx] if point_labels_batch is not None else None
+            )
+            box = box_batch[img_idx] if box_batch is not None else None
+            mask_input = (
+                mask_input_batch[img_idx] if mask_input_batch is not None else None
+            )
+            mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+                point_coords,
+                point_labels,
+                box,
+                mask_input,
+                normalize_coords,
+                img_idx=img_idx,
+            )
+            masks, iou_predictions, low_res_masks = self._predict(
+                unnorm_coords,
+                labels,
+                unnorm_box,
+                mask_input,
+                multimask_output,
+                return_logits=return_logits,
+                img_idx=img_idx,
+            )
+            masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+            iou_predictions_np = (
+                iou_predictions.squeeze(0).float().detach().cpu().numpy()
+            )
+            low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+            all_masks.append(masks_np)
+            all_ious.append(iou_predictions_np)
+            all_low_res_masks.append(low_res_masks_np)
+
+        return all_masks, all_ious, all_low_res_masks
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+          normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+
+        mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+            point_coords, point_labels, box, mask_input, normalize_coords
+        )
+
+        masks, iou_predictions, low_res_masks = self._predict(
+            unnorm_coords,
+            labels,
+            unnorm_box,
+            mask_input,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+        iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
+        low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    def _prep_prompts(
+        self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
+    ):
+
+        unnorm_coords, labels, unnorm_box, mask_input = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            unnorm_coords = self._transforms.transform_coords(
+                point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )
+            labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            if len(unnorm_coords.shape) == 2:
+                unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
+        if box is not None:
+            box = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            unnorm_box = self._transforms.transform_boxes(
+                box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )  # Bx2x2
+        if mask_logits is not None:
+            mask_input = torch.as_tensor(
+                mask_logits, dtype=torch.float, device=self.device
+            )
+            if len(mask_input.shape) == 3:
+                mask_input = mask_input[None, :, :, :]
+        return mask_input, unnorm_coords, labels, unnorm_box
+
+    @torch.no_grad()
+    def _predict(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        img_idx: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using SAM2Transforms.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            concat_points = (point_coords, point_labels)
+        else:
+            concat_points = None
+
+        # Embed prompts
+        if boxes is not None:
+            box_coords = boxes.reshape(-1, 2, 2)
+            box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
+            box_labels = box_labels.repeat(boxes.size(0), 1)
+            # we merge "boxes" and "points" into a single "concat_points" input (where
+            # boxes are added at the beginning) to sam_prompt_encoder
+            if concat_points is not None:
+                concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
+                concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
+                concat_points = (concat_coords, concat_labels)
+            else:
+                concat_points = (box_coords, box_labels)
+
+        sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
+            points=concat_points,
+            boxes=None,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        batched_mode = (
+            concat_points is not None and concat_points[0].shape[0] > 1
+        )  # multi object prediction
+        high_res_features = [
+            feat_level[img_idx].unsqueeze(0)
+            for feat_level in self._features["high_res_feats"]
+        ]
+        low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
+            image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
+            image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=batched_mode,
+            high_res_features=high_res_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self._transforms.postprocess_masks(
+            low_res_masks, self._orig_hw[img_idx]
+        )
+        low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
+        if not return_logits:
+            masks = masks > self.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self._features is not None
+        ), "Features must exist if an image has been set."
+        return self._features["image_embed"]
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_predictor(self) -> None:
+        """
+        Resets the image embeddings and other state variables.
+        """
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        self._is_batch = False

sam2/sam2_video_predictor.py

@@ -0,0 +1,957 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import warnings
+from collections import OrderedDict
+
+import torch
+
+from tqdm import tqdm
+
+from sam2.modeling.sam2_base import NO_OBJ_SCORE, SAM2Base
+from sam2.utils.misc import concat_points, fill_holes_in_mask_scores, load_video_frames
+
+
+class SAM2VideoPredictor(SAM2Base):
+    """The predictor class to handle user interactions and manage inference states."""
+
+    def __init__(
+        self,
+        fill_hole_area=0,
+        # whether to apply non-overlapping constraints on the output object masks
+        non_overlap_masks=False,
+        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
+        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
+        clear_non_cond_mem_around_input=False,
+        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
+        clear_non_cond_mem_for_multi_obj=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.fill_hole_area = fill_hole_area
+        self.non_overlap_masks = non_overlap_masks
+        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
+        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj
+
+    @torch.inference_mode()
+    def init_state(
+        self,
+        frame_paths,
+        offload_video_to_cpu=False,
+        offload_state_to_cpu=False,
+        async_loading_frames=False,
+    ):
+        """Initialize a inference state."""
+        images, video_height, video_width = load_video_frames(
+            img_paths=frame_paths,
+            image_size=self.image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            async_loading_frames=async_loading_frames,
+        )
+        inference_state = {}
+        inference_state["images"] = images
+        inference_state["num_frames"] = len(images)
+        # whether to offload the video frames to CPU memory
+        # turning on this option saves the GPU memory with only a very small overhead
+        inference_state["offload_video_to_cpu"] = offload_video_to_cpu
+        # whether to offload the inference state to CPU memory
+        # turning on this option saves the GPU memory at the cost of a lower tracking fps
+        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
+        # and from 24 to 21 when tracking two objects)
+        inference_state["offload_state_to_cpu"] = offload_state_to_cpu
+        # the original video height and width, used for resizing final output scores
+        inference_state["video_height"] = video_height
+        inference_state["video_width"] = video_width
+        inference_state["device"] = torch.device("cuda")
+        if offload_state_to_cpu:
+            inference_state["storage_device"] = torch.device("cpu")
+        else:
+            inference_state["storage_device"] = torch.device("cuda")
+        # inputs on each frame
+        inference_state["point_inputs_per_obj"] = {}
+        inference_state["mask_inputs_per_obj"] = {}
+        # visual features on a small number of recently visited frames for quick interactions
+        inference_state["cached_features"] = {}
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # mapping between client-side object id and model-side object index
+        inference_state["obj_id_to_idx"] = OrderedDict()
+        inference_state["obj_idx_to_id"] = OrderedDict()
+        inference_state["obj_ids"] = []
+        # A storage to hold the model's tracking results and states on each frame
+        inference_state["output_dict"] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
+        inference_state["output_dict_per_obj"] = {}
+        # A temporary storage to hold new outputs when user interact with a frame
+        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
+        inference_state["temp_output_dict_per_obj"] = {}
+        # Frames that already holds consolidated outputs from click or mask inputs
+        # (we directly use their consolidated outputs during tracking)
+        inference_state["consolidated_frame_inds"] = {
+            "cond_frame_outputs": set(),  # set containing frame indices
+            "non_cond_frame_outputs": set(),  # set containing frame indices
+        }
+        # metadata for each tracking frame (e.g. which direction it's tracked)
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"] = {}
+        # Warm up the visual backbone and cache the image feature on frame 0
+        self._get_image_feature(inference_state, frame_idx=0, batch_size=1)
+        return inference_state
+
+    @classmethod
+    def from_pretrained(cls, model_id: str, **kwargs) -> "SAM2VideoPredictor":
+        """
+        Load a pretrained model from the Hugging Face hub.
+
+        Arguments:
+          model_id (str): The Hugging Face repository ID.
+          **kwargs: Additional arguments to pass to the model constructor.
+
+        Returns:
+          (SAM2VideoPredictor): The loaded model.
+        """
+        from sam2.build_sam import build_sam2_video_predictor_hf
+
+        sam_model = build_sam2_video_predictor_hf(model_id, **kwargs)
+        return cls(sam_model)
+
+    def _obj_id_to_idx(self, inference_state, obj_id):
+        """Map client-side object id to model-side object index."""
+        obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
+        if obj_idx is not None:
+            return obj_idx
+
+        # This is a new object id not sent to the server before. We only allow adding
+        # new objects *before* the tracking starts.
+        allow_new_object = not inference_state["tracking_has_started"]
+        if allow_new_object:
+            # get the next object slot
+            obj_idx = len(inference_state["obj_id_to_idx"])
+            inference_state["obj_id_to_idx"][obj_id] = obj_idx
+            inference_state["obj_idx_to_id"][obj_idx] = obj_id
+            inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
+            # set up input and output structures for this object
+            inference_state["point_inputs_per_obj"][obj_idx] = {}
+            inference_state["mask_inputs_per_obj"][obj_idx] = {}
+            inference_state["output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            inference_state["temp_output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            return obj_idx
+        else:
+            raise RuntimeError(
+                f"Cannot add new object id {obj_id} after tracking starts. "
+                f"All existing object ids: {inference_state['obj_ids']}. "
+                f"Please call 'reset_state' to restart from scratch."
+            )
+
+    def _obj_idx_to_id(self, inference_state, obj_idx):
+        """Map model-side object index to client-side object id."""
+        return inference_state["obj_idx_to_id"][obj_idx]
+
+    def _get_obj_num(self, inference_state):
+        """Get the total number of unique object ids received so far in this session."""
+        return len(inference_state["obj_idx_to_id"])
+
+    @torch.inference_mode()
+    def add_new_points_or_box(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        points=None,
+        labels=None,
+        clear_old_points=True,
+        normalize_coords=True,
+        box=None,
+    ):
+        """Add new points to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if (points is not None) != (labels is not None):
+            raise ValueError("points and labels must be provided together")
+        if points is None and box is None:
+            raise ValueError("at least one of points or box must be provided as input")
+
+        if points is None:
+            points = torch.zeros(0, 2, dtype=torch.float32)
+        elif not isinstance(points, torch.Tensor):
+            points = torch.tensor(points, dtype=torch.float32)
+        if labels is None:
+            labels = torch.zeros(0, dtype=torch.int32)
+        elif not isinstance(labels, torch.Tensor):
+            labels = torch.tensor(labels, dtype=torch.int32)
+        if points.dim() == 2:
+            points = points.unsqueeze(0)  # add batch dimension
+        if labels.dim() == 1:
+            labels = labels.unsqueeze(0)  # add batch dimension
+
+        # If `box` is provided, we add it as the first two points with labels 2 and 3
+        # along with the user-provided points (consistent with how SAM 2 is trained).
+        if box is not None:
+            if not clear_old_points:
+                raise ValueError(
+                    "cannot add box without clearing old points, since "
+                    "box prompt must be provided before any point prompt "
+                    "(please use clear_old_points=True instead)"
+                )
+            if inference_state["tracking_has_started"]:
+                warnings.warn(
+                    "You are adding a box after tracking starts. SAM 2 may not always be "
+                    "able to incorporate a box prompt for *refinement*. If you intend to "
+                    "use box prompt as an *initial* input before tracking, please call "
+                    "'reset_state' on the inference state to restart from scratch.",
+                    category=UserWarning,
+                    stacklevel=2,
+                )
+            if not isinstance(box, torch.Tensor):
+                box = torch.tensor(box, dtype=torch.float32, device=points.device)
+            box_coords = box.reshape(1, 2, 2)
+            box_labels = torch.tensor([2, 3], dtype=torch.int32, device=labels.device)
+            box_labels = box_labels.reshape(1, 2)
+            points = torch.cat([box_coords, points], dim=1)
+            labels = torch.cat([box_labels, labels], dim=1)
+
+        if normalize_coords:
+            video_H = inference_state["video_height"]
+            video_W = inference_state["video_width"]
+            points = points / torch.tensor([video_W, video_H]).to(points.device)
+        # scale the (normalized) coordinates by the model's internal image size
+        points = points * self.image_size
+        points = points.to(inference_state["device"])
+        labels = labels.to(inference_state["device"])
+
+        if not clear_old_points:
+            point_inputs = point_inputs_per_frame.get(frame_idx, None)
+        else:
+            point_inputs = None
+        point_inputs = concat_points(point_inputs, points, labels)
+
+        point_inputs_per_frame[frame_idx] = point_inputs
+        mask_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        # Get any previously predicted mask logits on this object and feed it along with
+        # the new clicks into the SAM mask decoder.
+        prev_sam_mask_logits = None
+        # lookup temporary output dict first, which contains the most recent output
+        # (if not found, then lookup conditioning and non-conditioning frame output)
+        prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
+        if prev_out is None:
+            prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
+            if prev_out is None:
+                prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
+
+        if prev_out is not None and prev_out["pred_masks"] is not None:
+            prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
+            # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
+            prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=point_inputs,
+            mask_inputs=None,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    def add_new_points(self, *args, **kwargs):
+        """Deprecated method. Please use `add_new_points_or_box` instead."""
+        return self.add_new_points_or_box(*args, **kwargs)
+
+    @torch.inference_mode()
+    def add_new_mask(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        mask,
+    ):
+        """Add new mask to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if not isinstance(mask, torch.Tensor):
+            mask = torch.tensor(mask, dtype=torch.bool)
+        assert mask.dim() == 2
+        mask_H, mask_W = mask.shape
+        mask_inputs_orig = mask[None, None]  # add batch and channel dimension
+        mask_inputs_orig = mask_inputs_orig.float().to(inference_state["device"])
+
+        # resize the mask if it doesn't match the model's image size
+        if mask_H != self.image_size or mask_W != self.image_size:
+            mask_inputs = torch.nn.functional.interpolate(
+                mask_inputs_orig,
+                size=(self.image_size, self.image_size),
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,  # use antialias for downsampling
+            )
+            mask_inputs = (mask_inputs >= 0.5).float()
+        else:
+            mask_inputs = mask_inputs_orig
+
+        mask_inputs_per_frame[frame_idx] = mask_inputs
+        point_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    def _get_orig_video_res_output(self, inference_state, any_res_masks):
+        """
+        Resize the object scores to the original video resolution (video_res_masks)
+        and apply non-overlapping constraints for final output.
+        """
+        device = inference_state["device"]
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        any_res_masks = any_res_masks.to(device, non_blocking=True)
+        if any_res_masks.shape[-2:] == (video_H, video_W):
+            video_res_masks = any_res_masks
+        else:
+            video_res_masks = torch.nn.functional.interpolate(
+                any_res_masks,
+                size=(video_H, video_W),
+                mode="bilinear",
+                align_corners=False,
+            )
+        if self.non_overlap_masks:
+            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
+        return any_res_masks, video_res_masks
+
+    def _consolidate_temp_output_across_obj(
+        self,
+        inference_state,
+        frame_idx,
+        is_cond,
+        run_mem_encoder,
+        consolidate_at_video_res=False,
+    ):
+        """
+        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
+        a frame into a single output for all objects, including
+        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
+           `output_dict_per_obj` for this frame) or leave them as placeholder values
+           (if they don't exist in `output_dict_per_obj` for this frame);
+        2) if specified, rerun memory encoder after apply non-overlapping constraints
+           on the object scores.
+        """
+        batch_size = self._get_obj_num(inference_state)
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+        # Optionally, we allow consolidating the temporary outputs at the original
+        # video resolution (to provide a better editing experience for mask prompts).
+        if consolidate_at_video_res:
+            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
+            consolidated_H = inference_state["video_height"]
+            consolidated_W = inference_state["video_width"]
+            consolidated_mask_key = "pred_masks_video_res"
+        else:
+            consolidated_H = consolidated_W = self.image_size // 4
+            consolidated_mask_key = "pred_masks"
+
+        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
+        # will be added when rerunning the memory encoder after applying non-overlapping
+        # constraints to object scores. Its "pred_masks" are prefilled with a large
+        # negative value (NO_OBJ_SCORE) to represent missing objects.
+        consolidated_out = {
+            "maskmem_features": None,
+            "maskmem_pos_enc": None,
+            consolidated_mask_key: torch.full(
+                size=(batch_size, 1, consolidated_H, consolidated_W),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["storage_device"],
+            ),
+            "obj_ptr": torch.full(
+                size=(batch_size, self.hidden_dim),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+        }
+        empty_mask_ptr = None
+        for obj_idx in range(batch_size):
+            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
+            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
+            # we fall back and look up its previous output in "output_dict_per_obj".
+            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
+            # "output_dict_per_obj" to find a previous output for this object.
+            if out is None:
+                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
+            if out is None:
+                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
+            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
+            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
+            # placeholder above) and set its object pointer to be a dummy pointer.
+            if out is None:
+                # Fill in dummy object pointers for those objects without any inputs or
+                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
+                # i.e. when we need to build the memory for tracking).
+                if run_mem_encoder:
+                    if empty_mask_ptr is None:
+                        empty_mask_ptr = self._get_empty_mask_ptr(
+                            inference_state, frame_idx
+                        )
+                    # fill object pointer with a dummy pointer (based on an empty mask)
+                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
+                continue
+            # Add the temporary object output mask to consolidated output mask
+            obj_mask = out["pred_masks"]
+            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
+            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
+            else:
+                # Resize first if temporary object mask has a different resolution
+                resized_obj_mask = torch.nn.functional.interpolate(
+                    obj_mask,
+                    size=consolidated_pred_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
+            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]
+
+        # Optionally, apply non-overlapping constraints on the consolidated scores
+        # and rerun the memory encoder
+        if run_mem_encoder:
+            device = inference_state["device"]
+            high_res_masks = torch.nn.functional.interpolate(
+                consolidated_out["pred_masks"].to(device, non_blocking=True),
+                size=(self.image_size, self.image_size),
+                mode="bilinear",
+                align_corners=False,
+            )
+            if self.non_overlap_masks_for_mem_enc:
+                high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
+            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
+                inference_state=inference_state,
+                frame_idx=frame_idx,
+                batch_size=batch_size,
+                high_res_masks=high_res_masks,
+                is_mask_from_pts=True,  # these frames are what the user interacted with
+            )
+            consolidated_out["maskmem_features"] = maskmem_features
+            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc
+
+        return consolidated_out
+
+    def _get_empty_mask_ptr(self, inference_state, frame_idx):
+        """Get a dummy object pointer based on an empty mask on the current frame."""
+        # A dummy (empty) mask with a single object
+        batch_size = 1
+        mask_inputs = torch.zeros(
+            (batch_size, 1, self.image_size, self.image_size),
+            dtype=torch.float32,
+            device=inference_state["device"],
+        )
+
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # Feed the empty mask and image feature above to get a dummy object pointer
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=True,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            output_dict={},
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=False,
+            run_mem_encoder=False,
+            prev_sam_mask_logits=None,
+        )
+        return current_out["obj_ptr"]
+
+    @torch.inference_mode()
+    def propagate_in_video_preflight(self, inference_state):
+        """Prepare inference_state and consolidate temporary outputs before tracking."""
+        # Tracking has started and we don't allow adding new objects until session is reset.
+        inference_state["tracking_has_started"] = True
+        batch_size = self._get_obj_num(inference_state)
+
+        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
+        # add them into "output_dict".
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        output_dict = inference_state["output_dict"]
+        # "consolidated_frame_inds" contains indices of those frames where consolidated
+        # temporary outputs have been added (either in this call or any previous calls
+        # to `propagate_in_video_preflight`).
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        for is_cond in [False, True]:
+            # Separately consolidate conditioning and non-conditioning temp outptus
+            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+            # Find all the frames that contain temporary outputs for any objects
+            # (these should be the frames that have just received clicks for mask inputs
+            # via `add_new_points_or_box` or `add_new_mask`)
+            temp_frame_inds = set()
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
+            consolidated_frame_inds[storage_key].update(temp_frame_inds)
+            # consolidate the temprary output across all objects on this frame
+            for frame_idx in temp_frame_inds:
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state, frame_idx, is_cond=is_cond, run_mem_encoder=True
+                )
+                # merge them into "output_dict" and also create per-object slices
+                output_dict[storage_key][frame_idx] = consolidated_out
+                self._add_output_per_object(
+                    inference_state, frame_idx, consolidated_out, storage_key
+                )
+                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+                )
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+
+            # clear temporary outputs in `temp_output_dict_per_obj`
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                obj_temp_output_dict[storage_key].clear()
+
+        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
+        # output on the same frame in "non_cond_frame_outputs"
+        for frame_idx in output_dict["cond_frame_outputs"]:
+            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for obj_output_dict in inference_state["output_dict_per_obj"].values():
+            for frame_idx in obj_output_dict["cond_frame_outputs"]:
+                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+            assert frame_idx in output_dict["cond_frame_outputs"]
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+
+        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
+        # with either points or mask inputs (which should be true under a correct workflow).
+        all_consolidated_frame_inds = (
+            consolidated_frame_inds["cond_frame_outputs"]
+            | consolidated_frame_inds["non_cond_frame_outputs"]
+        )
+        input_frames_inds = set()
+        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
+            input_frames_inds.update(point_inputs_per_frame.keys())
+        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
+            input_frames_inds.update(mask_inputs_per_frame.keys())
+        assert all_consolidated_frame_inds == input_frames_inds
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        """Propagate the input points across frames to track in the entire video."""
+        self.propagate_in_video_preflight(inference_state)
+
+        output_dict = inference_state["output_dict"]
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        obj_ids = inference_state["obj_ids"]
+        num_frames = inference_state["num_frames"]
+        batch_size = self._get_obj_num(inference_state)
+        if len(output_dict["cond_frame_outputs"]) == 0:
+            raise RuntimeError("No points are provided; please add points first")
+        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+        )
+
+        # set start index, end index, and processing order
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(output_dict["cond_frame_outputs"])
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
+            if start_frame_idx > 0:
+                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
+            else:
+                processing_order = []  # skip reverse tracking if starting from frame 0
+        else:
+            end_frame_idx = min(
+                start_frame_idx + max_frame_num_to_track, num_frames - 1
+            )
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+
+        for frame_idx in tqdm(processing_order, desc="propagate in video"):
+            # We skip those frames already in consolidated outputs (these are frames
+            # that received input clicks or mask). Note that we cannot directly run
+            # batched forward on them via `_run_single_frame_inference` because the
+            # number of clicks on each object might be different.
+            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+                storage_key = "cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
+                storage_key = "non_cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+            else:
+                storage_key = "non_cond_frame_outputs"
+                current_out, pred_masks = self._run_single_frame_inference(
+                    inference_state=inference_state,
+                    output_dict=output_dict,
+                    frame_idx=frame_idx,
+                    batch_size=batch_size,
+                    is_init_cond_frame=False,
+                    point_inputs=None,
+                    mask_inputs=None,
+                    reverse=reverse,
+                    run_mem_encoder=True,
+                )
+                output_dict[storage_key][frame_idx] = current_out
+            # Create slices of per-object outputs for subsequent interaction with each
+            # individual object after tracking.
+            self._add_output_per_object(
+                inference_state, frame_idx, current_out, storage_key
+            )
+            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}
+
+            # Resize the output mask to the original video resolution (we directly use
+            # the mask scores on GPU for output to avoid any CPU conversion in between)
+            _, video_res_masks = self._get_orig_video_res_output(
+                inference_state, pred_masks
+            )
+            yield frame_idx, obj_ids, video_res_masks
+
+    def _add_output_per_object(
+        self, inference_state, frame_idx, current_out, storage_key
+    ):
+        """
+        Split a multi-object output into per-object output slices and add them into
+        `output_dict_per_obj`. The resulting slices share the same tensor storage.
+        """
+        maskmem_features = current_out["maskmem_features"]
+        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)
+
+        maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)
+
+        output_dict_per_obj = inference_state["output_dict_per_obj"]
+        for obj_idx, obj_output_dict in output_dict_per_obj.items():
+            obj_slice = slice(obj_idx, obj_idx + 1)
+            obj_out = {
+                "maskmem_features": None,
+                "maskmem_pos_enc": None,
+                "pred_masks": current_out["pred_masks"][obj_slice],
+                "obj_ptr": current_out["obj_ptr"][obj_slice],
+            }
+            if maskmem_features is not None:
+                obj_out["maskmem_features"] = maskmem_features[obj_slice]
+            if maskmem_pos_enc is not None:
+                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
+            obj_output_dict[storage_key][frame_idx] = obj_out
+
+    @torch.inference_mode()
+    def reset_state(self, inference_state):
+        """Remove all input points or mask in all frames throughout the video."""
+        self._reset_tracking_results(inference_state)
+        # Remove all object ids
+        inference_state["obj_id_to_idx"].clear()
+        inference_state["obj_idx_to_id"].clear()
+        inference_state["obj_ids"].clear()
+        inference_state["point_inputs_per_obj"].clear()
+        inference_state["mask_inputs_per_obj"].clear()
+        inference_state["output_dict_per_obj"].clear()
+        inference_state["temp_output_dict_per_obj"].clear()
+
+    def _reset_tracking_results(self, inference_state):
+        """Reset all tracking inputs and results across the videos."""
+        for v in inference_state["point_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["mask_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        for v in inference_state["temp_output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        inference_state["output_dict"]["cond_frame_outputs"].clear()
+        inference_state["output_dict"]["non_cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["non_cond_frame_outputs"].clear()
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"].clear()
+
+    def _get_image_feature(self, inference_state, frame_idx, batch_size):
+        """Compute the image features on a given frame."""
+        # Look up in the cache first
+        image, backbone_out = inference_state["cached_features"].get(
+            frame_idx, (None, None)
+        )
+        if backbone_out is None:
+            # Cache miss -- we will run inference on a single image
+            image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
+            backbone_out = self.forward_image(image)
+            # Cache the most recent frame's feature (for repeated interactions with
+            # a frame; we can use an LRU cache for more frames in the future).
+            inference_state["cached_features"] = {frame_idx: (image, backbone_out)}
+
+        # expand the features to have the same dimension as the number of objects
+        expanded_image = image.expand(batch_size, -1, -1, -1)
+        expanded_backbone_out = {
+            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
+            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
+        }
+        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
+            expanded_backbone_out["backbone_fpn"][i] = feat.expand(
+                batch_size, -1, -1, -1
+            )
+        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
+            pos = pos.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["vision_pos_enc"][i] = pos
+
+        features = self._prepare_backbone_features(expanded_backbone_out)
+        features = (expanded_image,) + features
+        return features
+
+    def _run_single_frame_inference(
+        self,
+        inference_state,
+        output_dict,
+        frame_idx,
+        batch_size,
+        is_init_cond_frame,
+        point_inputs,
+        mask_inputs,
+        reverse,
+        run_mem_encoder,
+        prev_sam_mask_logits=None,
+    ):
+        """Run tracking on a single frame based on current inputs and previous memory."""
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # point and mask should not appear as input simultaneously on the same frame
+        assert point_inputs is None or mask_inputs is None
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=is_init_cond_frame,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            output_dict=output_dict,
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=reverse,
+            run_mem_encoder=run_mem_encoder,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = current_out["maskmem_features"]
+        if maskmem_features is not None:
+            maskmem_features = maskmem_features.to(torch.bfloat16)
+            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        pred_masks_gpu = current_out["pred_masks"]
+        # potentially fill holes in the predicted masks
+        if self.fill_hole_area > 0:
+            pred_masks_gpu = fill_holes_in_mask_scores(
+                pred_masks_gpu, self.fill_hole_area
+            )
+        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(inference_state, current_out)
+        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
+        obj_ptr = current_out["obj_ptr"]
+        # make a compact version of this frame's output to reduce the state size
+        compact_current_out = {
+            "maskmem_features": maskmem_features,
+            "maskmem_pos_enc": maskmem_pos_enc,
+            "pred_masks": pred_masks,
+            "obj_ptr": obj_ptr,
+        }
+        return compact_current_out, pred_masks_gpu
+
+    def _run_memory_encoder(
+        self, inference_state, frame_idx, batch_size, high_res_masks, is_mask_from_pts
+    ):
+        """
+        Run the memory encoder on `high_res_masks`. This is usually after applying
+        non-overlapping constraints to object scores. Since their scores changed, their
+        memory also need to be computed again with the memory encoder.
+        """
+        # Retrieve correct image features
+        _, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
+            inference_state, frame_idx, batch_size
+        )
+        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            current_vision_feats=current_vision_feats,
+            feat_sizes=feat_sizes,
+            pred_masks_high_res=high_res_masks,
+            is_mask_from_pts=is_mask_from_pts,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = maskmem_features.to(torch.bfloat16)
+        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(
+            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
+        )
+        return maskmem_features, maskmem_pos_enc
+
+    def _get_maskmem_pos_enc(self, inference_state, current_out):
+        """
+        `maskmem_pos_enc` is the same across frames and objects, so we cache it as
+        a constant in the inference session to reduce session storage size.
+        """
+        model_constants = inference_state["constants"]
+        # "out_maskmem_pos_enc" should be either a list of tensors or None
+        out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        if out_maskmem_pos_enc is not None:
+            if "maskmem_pos_enc" not in model_constants:
+                assert isinstance(out_maskmem_pos_enc, list)
+                # only take the slice for one object, since it's same across objects
+                maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
+                model_constants["maskmem_pos_enc"] = maskmem_pos_enc
+            else:
+                maskmem_pos_enc = model_constants["maskmem_pos_enc"]
+            # expand the cached maskmem_pos_enc to the actual batch size
+            batch_size = out_maskmem_pos_enc[0].size(0)
+            expanded_maskmem_pos_enc = [
+                x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
+            ]
+        else:
+            expanded_maskmem_pos_enc = None
+        return expanded_maskmem_pos_enc
+
+    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
+        """
+        Remove the non-conditioning memory around the input frame. When users provide
+        correction clicks, the surrounding frames' non-conditioning memories can still
+        contain outdated object appearance information and could confuse the model.
+
+        This method clears those non-conditioning memories surrounding the interacted
+        frame to avoid giving the model both old and new information about the object.
+        """
+        r = self.memory_temporal_stride_for_eval
+        frame_idx_begin = frame_idx - r * self.num_maskmem
+        frame_idx_end = frame_idx + r * self.num_maskmem
+        output_dict = inference_state["output_dict"]
+        non_cond_frame_outputs = output_dict["non_cond_frame_outputs"]
+        for t in range(frame_idx_begin, frame_idx_end + 1):
+            non_cond_frame_outputs.pop(t, None)
+            for obj_output_dict in inference_state["output_dict_per_obj"].values():
+                obj_output_dict["non_cond_frame_outputs"].pop(t, None)

sam2/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam2/utils/amg.py

@@ -0,0 +1,348 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from copy import deepcopy
+from itertools import product
+from typing import Any, Dict, Generator, ItemsView, List, Tuple
+
+import numpy as np
+import torch
+
+# Very lightly adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/utils/amg.py
+
+
+class MaskData:
+    """
+    A structure for storing masks and their related data in batched format.
+    Implements basic filtering and concatenation.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        for v in kwargs.values():
+            assert isinstance(
+                v, (list, np.ndarray, torch.Tensor)
+            ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats = dict(**kwargs)
+
+    def __setitem__(self, key: str, item: Any) -> None:
+        assert isinstance(
+            item, (list, np.ndarray, torch.Tensor)
+        ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats[key] = item
+
+    def __delitem__(self, key: str) -> None:
+        del self._stats[key]
+
+    def __getitem__(self, key: str) -> Any:
+        return self._stats[key]
+
+    def items(self) -> ItemsView[str, Any]:
+        return self._stats.items()
+
+    def filter(self, keep: torch.Tensor) -> None:
+        for k, v in self._stats.items():
+            if v is None:
+                self._stats[k] = None
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = v[keep.detach().cpu().numpy()]
+            elif isinstance(v, list) and keep.dtype == torch.bool:
+                self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
+            elif isinstance(v, list):
+                self._stats[k] = [v[i] for i in keep]
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def cat(self, new_stats: "MaskData") -> None:
+        for k, v in new_stats.items():
+            if k not in self._stats or self._stats[k] is None:
+                self._stats[k] = deepcopy(v)
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = torch.cat([self._stats[k], v], dim=0)
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
+            elif isinstance(v, list):
+                self._stats[k] = self._stats[k] + deepcopy(v)
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def to_numpy(self) -> None:
+        for k, v in self._stats.items():
+            if isinstance(v, torch.Tensor):
+                self._stats[k] = v.float().detach().cpu().numpy()
+
+
+def is_box_near_crop_edge(
+    boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
+) -> torch.Tensor:
+    """Filter masks at the edge of a crop, but not at the edge of the original image."""
+    crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
+    orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
+    boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
+    near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
+    near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
+    near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
+    return torch.any(near_crop_edge, dim=1)
+
+
+def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
+    box_xywh = deepcopy(box_xyxy)
+    box_xywh[2] = box_xywh[2] - box_xywh[0]
+    box_xywh[3] = box_xywh[3] - box_xywh[1]
+    return box_xywh
+
+
+def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
+    assert len(args) > 0 and all(
+        len(a) == len(args[0]) for a in args
+    ), "Batched iteration must have inputs of all the same size."
+    n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
+    for b in range(n_batches):
+        yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
+
+
+def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
+    """
+    Encodes masks to an uncompressed RLE, in the format expected by
+    pycoco tools.
+    """
+    # Put in fortran order and flatten h,w
+    b, h, w = tensor.shape
+    tensor = tensor.permute(0, 2, 1).flatten(1)
+
+    # Compute change indices
+    diff = tensor[:, 1:] ^ tensor[:, :-1]
+    change_indices = diff.nonzero()
+
+    # Encode run length
+    out = []
+    for i in range(b):
+        cur_idxs = change_indices[change_indices[:, 0] == i, 1]
+        cur_idxs = torch.cat(
+            [
+                torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
+                cur_idxs + 1,
+                torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
+            ]
+        )
+        btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
+        counts = [] if tensor[i, 0] == 0 else [0]
+        counts.extend(btw_idxs.detach().cpu().tolist())
+        out.append({"size": [h, w], "counts": counts})
+    return out
+
+
+def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
+    """Compute a binary mask from an uncompressed RLE."""
+    h, w = rle["size"]
+    mask = np.empty(h * w, dtype=bool)
+    idx = 0
+    parity = False
+    for count in rle["counts"]:
+        mask[idx : idx + count] = parity
+        idx += count
+        parity ^= True
+    mask = mask.reshape(w, h)
+    return mask.transpose()  # Put in C order
+
+
+def area_from_rle(rle: Dict[str, Any]) -> int:
+    return sum(rle["counts"][1::2])
+
+
+def calculate_stability_score(
+    masks: torch.Tensor, mask_threshold: float, threshold_offset: float
+) -> torch.Tensor:
+    """
+    Computes the stability score for a batch of masks. The stability
+    score is the IoU between the binary masks obtained by thresholding
+    the predicted mask logits at high and low values.
+    """
+    # One mask is always contained inside the other.
+    # Save memory by preventing unnecessary cast to torch.int64
+    intersections = (
+        (masks > (mask_threshold + threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    unions = (
+        (masks > (mask_threshold - threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    return intersections / unions
+
+
+def build_point_grid(n_per_side: int) -> np.ndarray:
+    """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
+    offset = 1 / (2 * n_per_side)
+    points_one_side = np.linspace(offset, 1 - offset, n_per_side)
+    points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
+    points_y = np.tile(points_one_side[:, None], (1, n_per_side))
+    points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
+    return points
+
+
+def build_all_layer_point_grids(
+    n_per_side: int, n_layers: int, scale_per_layer: int
+) -> List[np.ndarray]:
+    """Generates point grids for all crop layers."""
+    points_by_layer = []
+    for i in range(n_layers + 1):
+        n_points = int(n_per_side / (scale_per_layer**i))
+        points_by_layer.append(build_point_grid(n_points))
+    return points_by_layer
+
+
+def generate_crop_boxes(
+    im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
+) -> Tuple[List[List[int]], List[int]]:
+    """
+    Generates a list of crop boxes of different sizes. Each layer
+    has (2**i)**2 boxes for the ith layer.
+    """
+    crop_boxes, layer_idxs = [], []
+    im_h, im_w = im_size
+    short_side = min(im_h, im_w)
+
+    # Original image
+    crop_boxes.append([0, 0, im_w, im_h])
+    layer_idxs.append(0)
+
+    def crop_len(orig_len, n_crops, overlap):
+        return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
+
+    for i_layer in range(n_layers):
+        n_crops_per_side = 2 ** (i_layer + 1)
+        overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
+
+        crop_w = crop_len(im_w, n_crops_per_side, overlap)
+        crop_h = crop_len(im_h, n_crops_per_side, overlap)
+
+        crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
+        crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
+
+        # Crops in XYWH format
+        for x0, y0 in product(crop_box_x0, crop_box_y0):
+            box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
+            crop_boxes.append(box)
+            layer_idxs.append(i_layer + 1)
+
+    return crop_boxes, layer_idxs
+
+
+def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
+    # Check if boxes has a channel dimension
+    if len(boxes.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return boxes + offset
+
+
+def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0]], device=points.device)
+    # Check if points has a channel dimension
+    if len(points.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return points + offset
+
+
+def uncrop_masks(
+    masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int
+) -> torch.Tensor:
+    x0, y0, x1, y1 = crop_box
+    if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
+        return masks
+    # Coordinate transform masks
+    pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
+    pad = (x0, pad_x - x0, y0, pad_y - y0)
+    return torch.nn.functional.pad(masks, pad, value=0)
+
+
+def remove_small_regions(
+    mask: np.ndarray, area_thresh: float, mode: str
+) -> Tuple[np.ndarray, bool]:
+    """
+    Removes small disconnected regions and holes in a mask. Returns the
+    mask and an indicator of if the mask has been modified.
+    """
+    import cv2  # type: ignore
+
+    assert mode in ["holes", "islands"]
+    correct_holes = mode == "holes"
+    working_mask = (correct_holes ^ mask).astype(np.uint8)
+    n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
+    sizes = stats[:, -1][1:]  # Row 0 is background label
+    small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
+    if len(small_regions) == 0:
+        return mask, False
+    fill_labels = [0] + small_regions
+    if not correct_holes:
+        fill_labels = [i for i in range(n_labels) if i not in fill_labels]
+        # If every region is below threshold, keep largest
+        if len(fill_labels) == 0:
+            fill_labels = [int(np.argmax(sizes)) + 1]
+    mask = np.isin(regions, fill_labels)
+    return mask, True
+
+
+def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
+    from pycocotools import mask as mask_utils  # type: ignore
+
+    h, w = uncompressed_rle["size"]
+    rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
+    rle["counts"] = rle["counts"].decode("utf-8")  # Necessary to serialize with json
+    return rle
+
+
+def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
+    an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
+    """
+    # torch.max below raises an error on empty inputs, just skip in this case
+    if torch.numel(masks) == 0:
+        return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
+
+    # Normalize shape to CxHxW
+    shape = masks.shape
+    h, w = shape[-2:]
+    if len(shape) > 2:
+        masks = masks.flatten(0, -3)
+    else:
+        masks = masks.unsqueeze(0)
+
+    # Get top and bottom edges
+    in_height, _ = torch.max(masks, dim=-1)
+    in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
+    bottom_edges, _ = torch.max(in_height_coords, dim=-1)
+    in_height_coords = in_height_coords + h * (~in_height)
+    top_edges, _ = torch.min(in_height_coords, dim=-1)
+
+    # Get left and right edges
+    in_width, _ = torch.max(masks, dim=-2)
+    in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
+    right_edges, _ = torch.max(in_width_coords, dim=-1)
+    in_width_coords = in_width_coords + w * (~in_width)
+    left_edges, _ = torch.min(in_width_coords, dim=-1)
+
+    # If the mask is empty the right edge will be to the left of the left edge.
+    # Replace these boxes with [0, 0, 0, 0]
+    empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
+    out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
+    out = out * (~empty_filter).unsqueeze(-1)
+
+    # Return to original shape
+    if len(shape) > 2:
+        out = out.reshape(*shape[:-2], 4)
+    else:
+        out = out[0]
+
+    return out

sam2/utils/misc.py

@@ -0,0 +1,238 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import warnings
+from threading import Thread
+
+import numpy as np
+import torch
+from PIL import Image
+from tqdm import tqdm
+
+
+def get_sdpa_settings():
+    if torch.cuda.is_available():
+        old_gpu = torch.cuda.get_device_properties(0).major < 7
+        # only use Flash Attention on Ampere (8.0) or newer GPUs
+        use_flash_attn = torch.cuda.get_device_properties(0).major >= 8
+        if not use_flash_attn:
+            warnings.warn(
+                "Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        # keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
+        # available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
+        pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
+        if pytorch_version < (2, 2):
+            warnings.warn(
+                f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
+                "Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
+    else:
+        old_gpu = True
+        use_flash_attn = False
+        math_kernel_on = True
+
+    return old_gpu, use_flash_attn, math_kernel_on
+
+
+def get_connected_components(mask):
+    """
+    Get the connected components (8-connectivity) of binary masks of shape (N, 1, H, W).
+
+    Inputs:
+    - mask: A binary mask tensor of shape (N, 1, H, W), where 1 is foreground and 0 is
+            background.
+
+    Outputs:
+    - labels: A tensor of shape (N, 1, H, W) containing the connected component labels
+              for foreground pixels and 0 for background pixels.
+    - counts: A tensor of shape (N, 1, H, W) containing the area of the connected
+              components for foreground pixels and 0 for background pixels.
+    """
+    from sam2 import csrc
+
+    return csrc.connect(mask.to(torch.uint8).contiguous())
+
+
+def mask_to_box(masks: torch.Tensor):
+    """
+    compute bounding box given an input mask
+
+    Inputs:
+    - masks: [B, 1, H, W] boxes, dtype=torch.Tensor
+
+    Returns:
+    - box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
+    """
+    B, _, h, w = masks.shape
+    device = masks.device
+    xs = torch.arange(w, device=device, dtype=torch.int32)
+    ys = torch.arange(h, device=device, dtype=torch.int32)
+    grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
+    grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
+    grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
+    min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
+    max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
+    min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
+    max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
+    bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
+
+    return bbox_coords
+
+
+def _load_img_as_tensor(img_path, image_size):
+    img_pil = Image.open(img_path)
+    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
+        img_np = img_np / 255.0
+    else:
+        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
+    img = torch.from_numpy(img_np).permute(2, 0, 1)
+    video_width, video_height = img_pil.size  # the original video size
+    return img, video_height, video_width
+
+
+class AsyncVideoFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(self, img_paths, image_size, offload_video_to_cpu, img_mean, img_std):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self._images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.cuda(non_blocking=True)
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+def load_video_frames(
+    img_paths,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+):
+    """
+    Load the video frames from a directory of JPEG/PNG files ("<frame_index>.jpg" or "<frame_index>.png" format).
+
+    The frames are resized to image_size x image_size and are loaded to GPU if
+    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.
+
+    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
+    """
+    num_frames = len(img_paths)
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncVideoFrameLoader(
+            img_paths, image_size, offload_video_to_cpu, img_mean, img_std
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
+    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def fill_holes_in_mask_scores(mask, max_area):
+    """
+    A post processor to fill small holes in mask scores with area under `max_area`.
+    """
+    # Holes are those connected components in background with area <= self.max_area
+    # (background regions are those with mask scores <= 0)
+    assert max_area > 0, "max_area must be positive"
+
+    input_mask = mask
+    try:
+        labels, areas = get_connected_components(mask <= 0)
+        is_hole = (labels > 0) & (areas <= max_area)
+        # We fill holes with a small positive mask score (0.1) to change them to foreground.
+        mask = torch.where(is_hole, 0.1, mask)
+    except Exception as e:
+        # Skip the post-processing step on removing small holes if the CUDA kernel fails
+        warnings.warn(
+            f"{e}\n\nSkipping the post-processing step due to the error above. "
+            "Consider building SAM 2 with CUDA extension to enable post-processing (see "
+            "https://github.com/facebookresearch/segment-anything-2/blob/main/INSTALL.md).",
+            category=UserWarning,
+            stacklevel=2,
+        )
+        mask = input_mask
+
+    return mask
+
+
+def concat_points(old_point_inputs, new_points, new_labels):
+    """Add new points and labels to previous point inputs (add at the end)."""
+    if old_point_inputs is None:
+        points, labels = new_points, new_labels
+    else:
+        points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
+        labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
+
+    return {"point_coords": points, "point_labels": labels}

sam2/utils/transforms.py

@@ -0,0 +1,117 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import warnings
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms import Normalize, Resize, ToTensor
+
+
+class SAM2Transforms(nn.Module):
+    def __init__(
+        self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
+    ):
+        """
+        Transforms for SAM2.
+        """
+        super().__init__()
+        self.resolution = resolution
+        self.mask_threshold = mask_threshold
+        self.max_hole_area = max_hole_area
+        self.max_sprinkle_area = max_sprinkle_area
+        self.mean = [0.485, 0.456, 0.406]
+        self.std = [0.229, 0.224, 0.225]
+        self.to_tensor = ToTensor()
+        self.transforms = torch.jit.script(
+            nn.Sequential(
+                Resize((self.resolution, self.resolution)),
+                Normalize(self.mean, self.std),
+            )
+        )
+
+    def __call__(self, x):
+        x = self.to_tensor(x)
+        return self.transforms(x)
+
+    def forward_batch(self, img_list):
+        img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
+        img_batch = torch.stack(img_batch, dim=0)
+        return img_batch
+
+    def transform_coords(
+        self, coords: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
+        If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+
+        Returns
+            Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
+        """
+        if normalize:
+            assert orig_hw is not None
+            h, w = orig_hw
+            coords = coords.clone()
+            coords[..., 0] = coords[..., 0] / w
+            coords[..., 1] = coords[..., 1] / h
+
+        coords = coords * self.resolution  # unnormalize coords
+        return coords
+
+    def transform_boxes(
+        self, boxes: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
+        if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+        """
+        boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
+        return boxes
+
+    def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
+        """
+        Perform PostProcessing on output masks.
+        """
+        from sam2.utils.misc import get_connected_components
+
+        masks = masks.float()
+        input_masks = masks
+        mask_flat = masks.flatten(0, 1).unsqueeze(1)  # flatten as 1-channel image
+        try:
+            if self.max_hole_area > 0:
+                # Holes are those connected components in background with area <= self.fill_hole_area
+                # (background regions are those with mask scores <= self.mask_threshold)
+                labels, areas = get_connected_components(
+                    mask_flat <= self.mask_threshold
+                )
+                is_hole = (labels > 0) & (areas <= self.max_hole_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with a small positive mask score (10.0) to change them to foreground.
+                masks = torch.where(is_hole, self.mask_threshold + 10.0, masks)
+
+            if self.max_sprinkle_area > 0:
+                labels, areas = get_connected_components(
+                    mask_flat > self.mask_threshold
+                )
+                is_hole = (labels > 0) & (areas <= self.max_sprinkle_area)
+                is_hole = is_hole.reshape_as(masks)
+                # We fill holes with negative mask score (-10.0) to change them to background.
+                masks = torch.where(is_hole, self.mask_threshold - 10.0, masks)
+        except Exception as e:
+            # Skip the post-processing step if the CUDA kernel fails
+            warnings.warn(
+                f"{e}\n\nSkipping the post-processing step due to the error above. "
+                "Consider building SAM 2 with CUDA extension to enable post-processing (see "
+                "https://github.com/facebookresearch/segment-anything-2/blob/main/INSTALL.md).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            masks = input_masks
+
+        masks = F.interpolate(masks, orig_hw, mode="bilinear", align_corners=False)
+        return masks

sam2_configs/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

sam2_configs/sam2_hiera_b+.yaml

@@ -0,0 +1,113 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 112
+      num_heads: 2
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [896, 448, 224, 112]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

sam2_configs/sam2_hiera_l.yaml

@@ -0,0 +1,117 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 144
+      num_heads: 2
+      stages: [2, 6, 36, 4]
+      global_att_blocks: [23, 33, 43]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+      window_spec: [8, 4, 16, 8]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [1152, 576, 288, 144]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

sam2_configs/sam2_hiera_s.yaml

@@ -0,0 +1,116 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 11, 2]
+      global_att_blocks: [7, 10, 13]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False

sam2_configs/sam2_hiera_t.yaml

@@ -0,0 +1,118 @@
+# @package _global_
+
+# Model
+model:
+  _target_: sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 7, 2]
+      global_att_blocks: [5, 7, 9]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  # SAM decoder
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  # HieraT does not currently support compilation, should always be set to False
+  compile_image_encoder: False

sam2point/configs copy.py

@@ -0,0 +1,363 @@
+sample_2 = {'path': 'data/S3DIS/Area_1_conferenceRoom_1.txt',
+           'point_prompts': [[0.01049672, 0.47400134, 0.51851852], [0.79906279, 0.88886409, 0.23477715], [0.62417994, 0.79825932, 0.01349655],
+                             [0.15126523, 0.88886409, 0.18047709], [0.54020619, 0.52041955, 0.24670433],],
+           'box_prompts': [[0.03, 0.63, 0.98, 0.18, 0.78, 1.0], [0.0, 0.4, 0.0, 0.15, 0.55, 0.27], [0.2, 0.95, 0.25, 0.7, 1.0, 0.67],
+                           [0.2, 0.2, 0.7, 0.25, 0.8, 0.78], [0.68, 0.85, 0., 1.0, 1.0, 0.25],[0, 0.82, 0.02, 0.2, 1, 0.38]],
+}
+
+
+sample_3 = {'path': 'data/S3DIS/Area_2_WC_1.txt',
+           'point_prompts': [[0.31414868, 0.59265659, 0.50951199], [0.6628697,  0.90842333, 0.34036394],[0.63868905, 0.36414687, 0.94954508],
+                             [0.11171063, 0.85788337, 0.18072787], [0.76159073, 0.82289417, 0.68899917],
+                             [0.88589129, 0.59049676, 0.44830438],],
+           'box_prompts': [[0.35, 0.8, 0.05, 0.45, 1.0, 0.4], [0.48, 0.65, 0.0, 0.55, 0.99, 0.99], [0.57, 0.2, 0.85, 0.7, 0.48, 1.0],
+                           [0.61, 0., 0.33, 0.71, 0.13, 0.51],], # [0.51, 0., 0., 0.61, 0.15, 0.37], 
+}
+
+
+sample_4 = {'path': 'data/S3DIS/Area_4_lobby_2.txt',
+           'point_prompts': [[0.19949431, 0.28597082, 0.25131625], [0.30316056, 0.87452301, 0.33696034],
+                             [0.72566372, 0.3617284,  0.65601966], [0.50316056, 0.57519641, 0.32186732],
+                             [0.46396966, 0.52345679, 0.54756055],],
+           'box_prompts': [[0.42, 0.45, 0.3, 0.49, 0.54, 0.65], [0.45, 0.57, 0.27, 0.55, 0.63, 0.36], [0.17, 0.35, 0., 0.25, 0.4, 0.3],
+                           [0.15, 0.25, 0.4, 0.19, 0.33, 0.62], [0.17, 0.78, 0.27, 0.2, 0.84, 0.43]],
+}
+
+sample_1 = {'path': 'data/S3DIS/Area_5_office_3.txt',
+           'point_prompts': [[0.55965254, 0.72432783, 0.00623636], [0.45080659, 0.88824101, 0.22856252],
+                             [0.90161319, 0.51668286, 0.21546617], [0.36589257, 0.93683188, 0.64826941], [0.98404538, 0.29024943, 0.51013408],
+                             [0.76369438, 0.32458698, 0.23542251]],
+           'box_prompts': [[0., 0.48, 0.23, 0.12, 0.61, 0.31], [0.4, 0.25, 0., 0.6, 0.6, 0.3], [0.45, 0.85, 0.45, 0.65, 0.99, 0.55],
+                           [0.38, 0.95, 0.25, 0.48, 1.00, 0.42], [0.65, 0.45, 0., 0.75, 0.6, 0.3]],
+}
+
+sample_0 = {'path': 'data/S3DIS/Area_6_office_9.txt',
+           'point_prompts': [[0.16548, 0.27853667, 0.1886402], [0.46150787, 0.09795895, 0.26989673], [0.2904479, 0.5073498,  0.28115318],
+                             [0.73819816, 0.913756, 0.2815835 ], [0.9304859, 0.40291342, 0.32013769], [0.802557, 0.5818576, 0.19074],
+                             [0.52659518, 0.5240772, 0.40165232], [0.29337714, 0.8905976, 0.2722375], [0.563984, 0.925, 0.3803788],
+                             [0.338812, 0.48102965, 0.34078142]],
+           'box_prompts': [[0.1, 0.2, 0.0, 0.2, 0.3, 0.4], [0.1, 0.02, 0.2, 0.9, 0.2, 0.3], [0.7, 0.5, 0., 0.9, 0.7, 0.4],
+                           [0.85, 0.3, 0.02, 0.98, 0.5, 0.8], [0.4, 0.4, 0.3, 0.6, 0.6, 0.5], ],
+}
+
+
+S3DIS_samples = [sample_2, sample_3, sample_4, sample_1, sample_0]
+
+
+
+
+sample_0 = {'path': 'data/ScanNet/scene0001_01.pth',
+           'point_prompts': [[0.48574361, 0.70011979, 0.21237852],
+                             [0.28947121, 0.15144145, 0.24688229], [0.3489365,  0.53977334, 0.02221746],
+                             [0.48059669, 0.88824904, 0.25690538]],  #[0.48760539, 0.12294616, 0.25476629],  #[0.48738128, 0.63986588, 0.25412986],
+           'box_prompts': [[0.25, 0.63, 0., 0.57, 0.75, 0.37], [0.42, 0.83, 0., 0.54, 0.94, 0.3], [0.4, 0.05, 0.0, 0.53, 0.2, 0.3],
+                           [0.12, 0.35, 0.0, 0.22, 0.45, 0.24], [0.88, 0.2, 0.1, 0.95, 0.8, 0.48]],
+}
+
+
+sample_1 = {'path': 'data/ScanNet/scene0005_01.pth', #[0.04293748, 0.38949549, 0.314679], [0.24069363, 0.51310396, 0.01414406], 
+           'point_prompts': [[0.50845712, 0.4027696,  0.19570725], [0.26778319, 0.9830749,  0.44313431]],  #[0.6458742,  0.33051795, 0.31433141], [0.11679079, 0.60943264, 0.40539789], 
+           'box_prompts': [[0.6, 0.6, 0., 0.83, 0.9, 0.33], [0.0, 0.57, 0.05, 0.15, 0.67, 0.48], #[0.41, 0.65, 0., 0.56, 0.77, 0.35],
+                           [0.48, 0.95, 0.58, 0.8, 0.99, 0.9]],
+}
+sample_2 = {'path': 'data/ScanNet/scene0010_01.pth',
+           'point_prompts': [[0.15311202, 0.44485098, 0.4582684], [0.86644632, 0.26297486, 0.5173167], [0.89919734, 0.40822271, 0.6298126 ]],  #,[0.66389197, 0.49352551, 0.2987611], [0.09592603, 0.20024474, 0.67744112]
+           'box_prompts': [[0.6, 0.72, 0.0, 0.75, 0.85, 0.6], [0.75, 0.75, 0.5, 0.92, 0.92, 0.75], [0.05, 0.92, 0.05, 0.27, 1.0, 0.82],
+                           [0.35, 0.03, 0.15, 0.5, 0.1, 0.42], ],
+}
+
+
+sample_3 = {'path': 'data/ScanNet/scene0016_02.pth',
+           'point_prompts': [[0.77345204, 0.5883323,  0.21049459], [0.82484114, 0.16314957, 0.23850442], [0.97325081, 0.28361404, 0.15121479],
+                             [0.29043797, 0.58934051, 0.82521498], [0.46316043, 0.34840286, 0.01032902], [0.3637068,  0.50896871, 0.63058698]],
+           'box_prompts': [[0.72, 0.36, 0.1, 0.9, 0.75, 0.75], [0.86, 0.12, 0.33, 0.99, 0.24, 0.54], [0.27, 0.54, 0.7, 0.3, 0.65, 0.9],
+                           [0.42, 0.5, 0.05, 0.55, 0.68, 0.42]],
+}
+
+
+sample_4 = {'path': 'data/ScanNet/scene0019_01.pth',
+           'point_prompts': [[0.52182293, 0.69650459, 0.36580974], [0.79430991, 0.31488013, 0.2448331], [0.6603151,  0.26341686, 0.33537653],
+                             [0.14427963, 0.69153076, 0.20673281], [0.17163187, 0.30585486, 0.31457961], [0.03188787, 0.65648252, 0.43863711]],
+           'box_prompts': [[0.55, 0.22, 0.05, 0.72, 0.3, 0.58], [0.0, 0.27, 0.05, 0.2, 0.35, 0.45], [0.03, 0.59, 0.05, 0.2, 0.85, 0.35],
+                           [0.43, 0.65, 0.05, 0.64, 0.72, 0.65]],
+}
+
+sample_5 = {'path': 'data/ScanNet/scene0000_00.pth',
+           'point_prompts': [[0.37658614, 0.11185088, 0.25310564], [0.40517676, 0.7643317,  0.16952564], [0.42705029, 0.8192997,  0.17624393]],
+           'box_prompts': [],
+}
+sample_6 = {'path': 'data/ScanNet/scene0002_00.pth',
+           'point_prompts': [[0.56711978, 0.74271345, 0.1753805 ], [0.61877084, 0.47617316, 0.23380645]],
+           'box_prompts': [],
+}
+
+ScanNet_samples = [sample_1, sample_2, sample_3, sample_4, sample_5, sample_6]  #sample_0,
+
+
+sample_0 = {'path': 'data/Objaverse/plant.npy',
+           'point_prompts': [[0.50455284, 0.47794762, 0.0007253083], [0.28331658, 0.19435011, 0.77393067]], #[7006, 1458],
+           'voxel_size': [0.038, 0.04],
+        #    'voxel_size': [0.03, 0.04],
+           'box_prompts': [[0.08, 0.18, -0.02, 0.68, 0.73, 0.315]], #, [0, 0, 0.3, 1, 1, 1.01]], #[0.11, 0.43, 0.82, 0.5, 1.01, 1.01]], 
+           'voxel_size_box': [0.04, 0.05], #0.01,
+           'mask_prompts': [[0.50455284, 0.47794762, 0.0007253083]], #[7006, 1458], , [0.28331658, 0.19435011, 0.77393067]
+           'voxel_size_mask': [0.038]
+}
+
+
+sample_1 = {'path': 'data/Objaverse/human.npy',
+           'point_prompts': [[0.57825595, 0.5005686,  0.11494722], [0.7136412,  0.49501216, 0.5020814 ], [0.7136412,  0.49501216, 0.5020814 ]], #[1112, 2133, 2133],
+           'voxel_size': [0.055, 0.045, 0.05],
+           'box_prompts': [[0., 0.17, -0.01, 0.72, 0.80, 0.3], [-0.01, 0., 0.28, 0.8, 1, 0.82], [-0.01, 0.28, 0.89, 1, 0.72, 1.02]],
+           'voxel_size_box': [0.055, 0.045, 0.055],
+           'mask_prompts': [[0.57825595, 0.5005686,  0.11494722], [0.7136412,  0.49501216, 0.5020814 ]], #[1112, 2133, 2133],
+           'voxel_size_mask': [0.055, 0.055],
+}
+sample_2 = {'path': 'data/Objaverse/lock.npy',
+           'point_prompts': [[0.6513301, 0.6753892, 0.52316076], [0.21359734, 0.6097132 , 0.7939796 ], [0.44947368, 0.21654338, 0.58450174]], #[1029, 2064, 3541],  #, [0.67447126, 0.6777649 , 0.51486933]
+           'voxel_size': [0.04, 0.05, 0.05], #, 0.05
+           'box_prompts': [[0.61, 0.4, 0.35, 0.8, 0.8, 0.6], [0.42, -0.02, -0.02, 1.02, 0.4, 1]],  #[0., 0.25, -0.02, 0.4, 0.82, 1],
+           'voxel_size_box': [0.04, 0.011],  # 0.05, 0.04
+           'mask_prompts': [[0.6513301, 0.6753892, 0.52316076], [0.21359734, 0.6097132 , 0.7939796 ], [0.9157764,  0.1995991,  0.14024617]], #[1029, 2064, 3541],
+           'voxel_size_mask': [0.04, 0.055, 0.04],
+}
+
+sample_3 = {'path': 'data/Objaverse/elephant.npy',
+           'point_prompts': [[0.4394578, 0.8342078, 0.835564]],
+           'voxel_size': [0.04],
+           'box_prompts': [[0.25,0,0,0.8,0.35,0.23]],
+           'voxel_size_box': [0.04],
+           'mask_prompts': [[0.4394578, 0.8342078, 0.835564]],
+           'voxel_size_mask': [0.04],
+}
+
+sample_4 = {'path': 'data/Objaverse/knife_rest.npy',
+           'point_prompts': [[0.3342131, 0.5378736, 0.8621972], [0.7043406, 0.4798344, 0.2585481]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.21, 0.26, 0.83, 0.37, 0.9, 1], [0, 0, 0, 1, 1, 0.28]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.3342131, 0.5378736, 0.8621972]],
+           'voxel_size_mask': [0.04],
+}
+
+sample_5 = {'path': 'data/Objaverse/skateboard.npy',
+           'point_prompts': [[0.5026503, 0.4316724, 0.5640968], [0.2835252, 0.4883442, 0.2073544]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0, 0, 0.54, 1, 1, 1], [0.21, 0.75, 0, 0.34, 1, 0.5]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.5026503, 0.4316724, 0.5640968], [0.2835252, 0.4883442, 0.2073544]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_6 = {'path': 'data/Objaverse/popcorn_machine.npy',
+           'point_prompts': [[0.278306, 0.4913014, 0.7318756], [0.5867118, 0.1180351, 0.5844101]], #, [0.8857, 0.8296, 0.6090]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.208, 0.157, 0.493, 0.779, 0.89, 0.925]],
+           'voxel_size_box': [0.04],
+           'mask_prompts': [[0.278306, 0.4913014, 0.7318756], [0.5867118, 0.1180351, 0.5844101]], #, [0.8857, 0.8296, 0.6090]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_7 = {'path': 'data/Objaverse/stove.npy',
+           'point_prompts': [[0.08, 0.72, 0.669], [0.9416, 0.3464, 0.3476], [0.021837, 0.281256, 0.8934]],
+           'voxel_size': [0.04, 0.04, 0.04],
+           'box_prompts': [[0,0,0.579,0.18,1,0.67], [0.528, 0.64, 0.508, 0.844, 0.866, 0.56]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.08, 0.72, 0.669], [0.9416, 0.3464, 0.3476], [0.021837, 0.281256, 0.8934]], 
+           'voxel_size_mask': [0.04, 0.04, 0.04],
+}
+
+
+sample_8 = {'path': 'data/Objaverse/bus_shelter.npy',
+           'point_prompts': [[0.6665938, 0.5713098, 0.2139242], [0.577489, 0.915092, 0.4498839]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.32, 0.36, 0, 0.924, 0.861, 0.394], [0, 0, 0.71, 1, 1, 1]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.6665938, 0.5713098, 0.2139242], [0.577489, 0.915092, 0.4498839]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_9 = {'path': 'data/Objaverse/thor_hammer.npy',
+           'point_prompts': [[0.6211515, 0.5109989, 0.3867725], [0.44443, 0.2363458, 0.7229376]],
+           'voxel_size': [0.05, 0.05, 0.05],
+           'box_prompts': [[0,0,0.723,1,1,1]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.05, 0.05],
+           'mask_prompts': [[0.44443, 0.2363458, 0.7229376]],
+           'voxel_size_mask': [0.05],
+}
+
+sample_10 = {'path': 'data/Objaverse/horse.npy',
+           'point_prompts': [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.65,0,0.3,1,1,0.79], [0.37, 0, 0, 1, 1, 0.2]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts':  [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_11 = {'path': 'data/Objaverse/dinner_booth.npy',
+           'point_prompts': [
+    [0.9192697, 0.4469184, 0.0017635],
+    [0.4987888, 0.6916906, 0.5106028]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.65,0,0.3,1,1,0.79], [0.37, 0, 0, 1, 1, 0.2]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts':  [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+# sculpture.npy
+# horse.npy
+# pipe.npy
+# dinner_booth.npy
+# ornament.npy
+# blender.npy
+# bowl.npy
+# human_face.npy
+# table.npy
+# telescope.npy
+# planet.npy
+# lamp.npy
+# dragon.npy
+
+Objaverse_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5, sample_6, sample_7, sample_8, sample_9, sample_10, sample_11]
+# sample_1, sample_2, 
+
+
+
+sample_0 = {'path': 'data/KITTI/scene1.npy',
+           'point_prompts': [[0.5527776, 0.7294311, 0.685305 ]],
+           'voxel_size': [0.02],
+           'box_prompts': [[0.52, 0.73, 0.56, 0.57, 0.76, 0.75]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5527776, 0.7294311, 0.685305 ]],
+           'voxel_size_mask': [0.02],
+}
+
+
+sample_1 = {'path': 'data/KITTI/scene2.npy',
+           'point_prompts': [[0.5090489, 0.45589063, 0.49851784]],
+           'voxel_size': [0.015],
+           'box_prompts': [[0.48, 0.43, 0.34, 0.54, 0.48, 0.71]],
+           'voxel_size_box': [0.015],
+           'mask_prompts': [[0.5090489, 0.45589063, 0.49851784]],
+           'voxel_size_mask': [0.015],
+}
+
+
+sample_2 = {'path': 'data/KITTI/scene3.npy',
+           'point_prompts': [[0.5442487, 0.5907391, 0.5992437]],
+           'voxel_size': [0.01],
+           'box_prompts': [[0.532, 0.58, 0.37, 0.555, 0.61, 0.68]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5442487, 0.5907391, 0.5992437]],
+           'voxel_size_mask': [0.01],
+}
+
+sample_3 = {'path': 'kitti/scene4.npy',
+           'point_prompts': [[0.4739189, 0.4791307, 0.8351399]],
+           'voxel_size': [0.01],
+           'box_prompts': [[0.51, 0.2, 0.75, 0.53, 0.22, 0.9]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.4739189, 0.4791307, 0.8351399], [0.4585995, 0.4209206, 0.7708794]],
+           'voxel_size_mask': [0.01, 0.006],
+}
+
+sample_4 = {'path': 'kitti/scene5.npy',
+           'point_prompts': [[0.5438917, 0.7608865, 0.5123742], [0.5131016, 0.7495122, 0.5516282]],
+           'voxel_size': [0.01, 0.01],
+           'box_prompts': [[0.43, 0.746, 0.39, 0.471,0.77, 0.62]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5438917, 0.7608865, 0.5123742], [0.5131016, 0.7495122, 0.5516282]],
+           'voxel_size_mask': [0.01, 0.01, 0.01],
+}
+
+sample_5 = {'path': 'kitti/scene6.npy',
+           'point_prompts': [[0.4619498, 0.3496694, 0.7484359], [0.4963415, 0.5221788, 0.7358279]],
+           'voxel_size': [0.008, 0.01],
+           'box_prompts': [[0.5459, 0.4, 0.62, 0.559, 0.5, 0.77], [0.61,0.343,0.625,0.664,0.377,0.8261]],
+           'voxel_size_box': [0.01, 0.01],
+           'mask_prompts': [[0.4619498, 0.3496694, 0.7484359], [0.4963415, 0.5221788, 0.7358279]],
+           'voxel_size_mask': [0.008, 0.01],
+}
+KITTI_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5]
+
+
+
+
+sample_0 = {'path': 'data/Semantic3D/scene1.npy',
+           'point_prompts': [[0.08373796, 0.61115538, 0.6007256], [0.2660193, 0.823606, 0.242315]],
+           'voxel_size': [0.017, 0.017],
+           'box_prompts': [[-0.02, 0.52, -0.02, 0.1, 0.7, 0.92]],
+           'voxel_size_box': [0.017],
+           'mask_prompts': [[0.08373796, 0.61115538, 0.6007256]],
+           'voxel_size_mask': [0.017],
+}
+
+
+sample_1 = {'path': 'data/Semantic3D/scene2.npy',
+           'point_prompts': [[0.79984724, 0.25791535, 0.18132911]],
+           'voxel_size': [0.012],
+           'box_prompts': [[0.78, 0, -0.02, 1, 0.5, 0.2]],
+           'voxel_size_box': [0.012],
+           'mask_prompts': [[0.79984724, 0.25791535, 0.18132911]],
+           'voxel_size_mask': [0.012],
+}
+
+
+
+sample_2 = {'path': 'data/Semantic3D/patch19.npy',
+           'point_prompts': [[0.51970197, 0.38389998, 0.33622117],
+                             [0.84013408, 0.80095002, 0.24210576]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [[0.51970197, 0.38389998, 0.33622117],
+                             [0.84013408, 0.80095002, 0.24210576]],
+           'voxel_size_mask': [0.017, 0.017],
+}
+
+sample_3 = {'path': 'data/Semantic3D/patch0.npy',
+           'point_prompts': [[0.91819174, 0.34150001, 0.25513778], [0., 0.34900001, 0.32881831]],
+           'voxel_size': [0.015, 0.017, 0.017, 0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+sample_4 = {'path': 'data/Semantic3D/patch1.npy',
+           'point_prompts': [[0.51603703, 0.51312565, 0.50598845]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [[0.1857393, 0.2675134, 0.2463012]], #[[0.51603703, 0.51312565, 0.50598845]],
+           'voxel_size_mask': [0.01], #[0.01],
+}
+
+sample_5 = {'path': 'data/Semantic3D/patch50.npy',
+           'point_prompts': [[0.22901525, 0.49448244, 0.52076028]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [[0.09, 0.44, 0.08, 0.4, 0.75, 0.98]],
+           'voxel_size_box': [0.017, 0.017],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+
+sample_6 = {'path': 'data/Semantic3D/patch62.npy',
+           'point_prompts': [],
+           'voxel_size': [],
+           'box_prompts': [[0.26, 0.38, 0.24, 0.55, 0.78, 0.99]],
+           'voxel_size_box': [0.017],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+Semantic3D_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5, sample_6]
+
+
+
+
+VOXEL = {"point": "voxel_size", "box": "voxel_size_box", "mask": "voxel_size_mask"}
+
+

sam2point/configs.py

@@ -0,0 +1,384 @@
+sample_2 = {'path': 'data/S3DIS/Area_1_conferenceRoom_1.txt',
+           'point_prompts': [[0.01049672, 0.47400134, 0.51851852], [0.79906279, 0.88886409, 0.23477715], [0.62417994, 0.79825932, 0.01349655],
+                             [0.15126523, 0.88886409, 0.18047709], [0.54020619, 0.52041955, 0.24670433],],
+           'box_prompts': [[0.03, 0.63, 0.98, 0.18, 0.78, 1.0], [0.0, 0.4, 0.0, 0.15, 0.55, 0.27], [0.2, 0.95, 0.25, 0.7, 1.0, 0.67],
+                           [0.2, 0.2, 0.7, 0.25, 0.8, 0.78], [0.68, 0.85, 0., 1.0, 1.0, 0.25],[0, 0.82, 0.02, 0.2, 1, 0.38]],
+           'mask_prompts': [[0.01049672, 0.47400134, 0.51851852], [0.79906279, 0.88886409, 0.23477715], [0.62417994, 0.79825932, 0.01349655],
+                             [0.15126523, 0.88886409, 0.18047709], [0.54020619, 0.52041955, 0.24670433],],
+}
+
+
+sample_3 = {'path': 'data/S3DIS/Area_2_WC_1.txt',
+           'point_prompts': [[0.31414868, 0.59265659, 0.50951199], [0.6628697,  0.90842333, 0.34036394],[0.63868905, 0.36414687, 0.94954508],
+                             [0.11171063, 0.85788337, 0.18072787], #[0.76159073, 0.82289417, 0.68899917],
+                             [0.88589129, 0.59049676, 0.44830438],],
+           'box_prompts': [[0.35, 0.8, 0.05, 0.45, 1.0, 0.4], [0.48, 0.65, 0.0, 0.55, 0.99, 0.99], [0.57, 0.2, 0.85, 0.7, 0.48, 1.0],
+                           [0.61, 0., 0.33, 0.71, 0.13, 0.51],], # [0.51, 0., 0., 0.61, 0.15, 0.37], 
+           'mask_prompts': [[0.31414868, 0.59265659, 0.50951199], [0.6628697,  0.90842333, 0.34036394],[0.63868905, 0.36414687, 0.94954508],
+                             [0.11171063, 0.85788337, 0.18072787], #[0.76159073, 0.82289417, 0.68899917],
+                             [0.88589129, 0.59049676, 0.44830438],],
+}
+
+
+sample_4 = {'path': 'data/S3DIS/Area_4_lobby_2.txt',
+           'point_prompts': [[0.19949431, 0.28597082, 0.25131625], #[0.30316056, 0.87452301, 0.33696034],
+                             [0.72566372, 0.3617284,  0.65601966], [0.50316056, 0.57519641, 0.32186732],
+                             [0.46396966, 0.52345679, 0.54756055],],
+           'box_prompts': [[0.42, 0.45, 0.3, 0.49, 0.54, 0.65], [0.45, 0.57, 0.27, 0.55, 0.63, 0.36], [0.17, 0.35, 0., 0.25, 0.4, 0.3],
+                           [0.15, 0.25, 0.4, 0.19, 0.33, 0.62], [0.17, 0.78, 0.27, 0.2, 0.84, 0.43]],
+           'mask_prompts': [#[0.19949431, 0.28597082, 0.25131625], [0.30316056, 0.87452301, 0.33696034],
+                             [0.72566372, 0.3617284,  0.65601966], [0.50316056, 0.57519641, 0.32186732],
+                             [0.46396966, 0.52345679, 0.54756055],],
+}
+
+sample_1 = {'path': 'data/S3DIS/Area_5_office_3.txt',
+           'point_prompts': [ #[0.45080659, 0.88824101, 0.22856252], [0.55965254, 0.72432783, 0.00623636], [0.36589257, 0.93683188, 0.64826941], 
+                             [0.90161319, 0.51668286, 0.21546617], [0.98404538, 0.29024943, 0.51013408],
+                             [0.76369438, 0.32458698, 0.23542251]],
+           'box_prompts': [[0., 0.48, 0.23, 0.12, 0.61, 0.31], [0.4, 0.25, 0., 0.6, 0.6, 0.3], [0.45, 0.85, 0.45, 0.65, 0.99, 0.55],
+                           [0.38, 0.95, 0.25, 0.48, 1.00, 0.42], [0.65, 0.45, 0., 0.75, 0.6, 0.3]],
+           'mask_prompts': [[0.45080659, 0.88824101, 0.22856252],
+                             [0.90161319, 0.51668286, 0.21546617], [0.98404538, 0.29024943, 0.51013408],
+                             [0.76369438, 0.32458698, 0.23542251]],  #[0.55965254, 0.72432783, 0.00623636], [0.36589257, 0.93683188, 0.64826941],
+}
+
+sample_0 = {'path': 'data/S3DIS/Area_6_office_9.txt',
+           'point_prompts': [[0.16548, 0.27853667, 0.1886402], [0.46150787, 0.09795895, 0.26989673], [0.2904479, 0.5073498,  0.28115318],
+                             [0.9304859, 0.40291342, 0.32013769], [0.802557, 0.5818576, 0.19074],
+                             [0.52659518, 0.5240772, 0.40165232], [0.29337714, 0.8905976, 0.2722375], [0.563984, 0.925, 0.3803788],],
+                            #  [0.73819816, 0.913756, 0.2815835 ], [0.338812, 0.48102965, 0.34078142]],
+           'box_prompts': [[0.1, 0.2, 0.0, 0.2, 0.3, 0.4], [0.1, 0.02, 0.2, 0.9, 0.2, 0.3], [0.7, 0.5, 0., 0.9, 0.7, 0.4],
+                           [0.85, 0.3, 0.02, 0.98, 0.5, 0.8], [0.4, 0.4, 0.3, 0.6, 0.6, 0.5], ],
+           'mask_prompts': [[0.16548, 0.27853667, 0.1886402], [0.46150787, 0.09795895, 0.26989673], [0.2904479, 0.5073498,  0.28115318],
+                             [0.9304859, 0.40291342, 0.32013769], [0.802557, 0.5818576, 0.19074],
+                             [0.52659518, 0.5240772, 0.40165232], [0.29337714, 0.8905976, 0.2722375], [0.563984, 0.925, 0.3803788],]
+                            #  [0.73819816, 0.913756, 0.2815835 ], [0.338812, 0.48102965, 0.34078142]],
+}
+
+
+S3DIS_samples = [sample_2, sample_3, sample_4, sample_1, sample_0]
+
+
+
+
+# sample_0 = {'path': 'data/ScanNet/scene0001_01.pth',
+#            'point_prompts': [[0.48574361, 0.70011979, 0.21237852],
+#                              [0.28947121, 0.15144145, 0.24688229], [0.3489365,  0.53977334, 0.02221746],
+#                              [0.48059669, 0.88824904, 0.25690538]],  #[0.48760539, 0.12294616, 0.25476629],  #[0.48738128, 0.63986588, 0.25412986],
+#            'box_prompts': [[0.25, 0.63, 0., 0.57, 0.75, 0.37], [0.42, 0.83, 0., 0.54, 0.94, 0.3], [0.4, 0.05, 0.0, 0.53, 0.2, 0.3],
+#                            [0.12, 0.35, 0.0, 0.22, 0.45, 0.24], [0.88, 0.2, 0.1, 0.95, 0.8, 0.48]],
+# }
+
+
+sample_1 = {'path': 'data/ScanNet/scene0005_01.pth', #[0.04293748, 0.38949549, 0.314679], [0.24069363, 0.51310396, 0.01414406], 
+           'point_prompts': [[0.50845712, 0.4027696,  0.19570725], [0.26778319, 0.9830749,  0.44313431]],  #[0.6458742,  0.33051795, 0.31433141], [0.11679079, 0.60943264, 0.40539789], 
+           'box_prompts': [[0.6, 0.6, 0., 0.83, 0.9, 0.33], [0.0, 0.57, 0.05, 0.15, 0.67, 0.48], #[0.41, 0.65, 0., 0.56, 0.77, 0.35],
+                           [0.48, 0.95, 0.58, 0.8, 0.99, 0.9]],
+           'mask_prompts': [[0.50845712, 0.4027696,  0.19570725], [0.26778319, 0.9830749,  0.44313431]],
+}
+sample_2 = {'path': 'data/ScanNet/scene0010_01.pth',
+           'point_prompts': [[0.86644632, 0.26297486, 0.5173167]], #[0.15311202, 0.44485098, 0.4582684], [0.89919734, 0.40822271, 0.6298126 ]],  #,[0.66389197, 0.49352551, 0.2987611], [0.09592603, 0.20024474, 0.67744112]
+           'box_prompts': [[0.6, 0.72, 0.0, 0.75, 0.85, 0.6], [0.75, 0.70, 0.5, 0.92, 0.92, 0.75], [0.05, 0.92, 0.05, 0.27, 1.0, 0.82],
+                           [0.35, 0.03, 0.15, 0.5, 0.1, 0.42], ],
+           'mask_prompts': [[0.86644632, 0.26297486, 0.5173167]],
+}
+
+
+sample_3 = {'path': 'data/ScanNet/scene0016_02.pth',
+           'point_prompts': [[0.2898192, 0.5845358, 0.7862434], [0.8251329,0.1763976,0.2942619]],
+       #     [[0.77345204, 0.5883323,  0.21049459], [0.82484114, 0.16314957, 0.23850442], [0.97325081, 0.28361404, 0.15121479],
+                            #  [0.29043797, 0.58934051, 0.82521498], [0.46316043, 0.34840286, 0.01032902], [0.3637068,  0.50896871, 0.63058698]],
+           'box_prompts': [[0.72, 0.36, 0.1, 0.9, 0.75, 0.75], [0.27, 0.54, 0.7, 0.3, 0.65, 0.9],], #[0.86, 0.12, 0.33, 0.99, 0.24, 0.54], [0.42, 0.5, 0.05, 0.55, 0.68, 0.42]
+           'mask_prompts': [[0.2898192, 0.5845358, 0.7862434]],
+}
+
+
+sample_4 = {'path': 'data/ScanNet/scene0019_01.pth',
+           'point_prompts': [[0.52182293, 0.69650459, 0.36580974], [0.6603151,  0.26341686, 0.33537653],[0.03188787, 0.65648252, 0.43863711]], #
+                            #  [0.79430991, 0.31488013, 0.2448331], [0.14427963, 0.69153076, 0.20673281], [0.17163187, 0.30585486, 0.31457961], 
+           'box_prompts': [[0.55, 0.22, 0.05, 0.72, 0.3, 0.58], [0.0, 0.27, 0.05, 0.2, 0.35, 0.45]], #[0.03, 0.59, 0.05, 0.2, 0.85, 0.35],
+                     #       [0.43, 0.65, 0.05, 0.64, 0.72, 0.65]],
+           'mask_prompts': [[0.52182293, 0.69650459, 0.36580974], [0.6603151,  0.26341686, 0.33537653], [0.17163187, 0.30585486, 0.31457961], [0.03188787, 0.65648252, 0.43863711]],
+}
+
+sample_5 = {'path': 'data/ScanNet/scene0000_00.pth',
+           'point_prompts': [[0.37658614, 0.11185088, 0.25310564], [0.40517676, 0.7643317,  0.16952564], [0.42705029, 0.8192997,  0.17624393]],
+           'box_prompts': [],
+           'mask_prompts': [[0.37658614, 0.11185088, 0.25310564], [0.42705029, 0.8192997,  0.17624393]],
+}
+sample_6 = {'path': 'data/ScanNet/scene0002_00.pth',
+           'point_prompts': [[0.56711978, 0.74271345, 0.1753805 ], [0.61877084, 0.47617316, 0.23380645]],
+           'box_prompts': [],
+           'mask_prompts': [[0.56711978, 0.74271345, 0.1753805 ], [0.61877084, 0.47617316, 0.23380645]],
+}
+
+ScanNet_samples = [sample_1, sample_2, sample_3, sample_4, sample_5, sample_6]  #sample_0,
+
+
+sample_0 = {'path': 'data/Objaverse/plant.npy',
+           'point_prompts': [[0.50455284, 0.47794762, 0.0007253083], [0.28331658, 0.19435011, 0.77393067]], #[7006, 1458],
+           'voxel_size': [0.038, 0.04],
+        #    'voxel_size': [0.03, 0.04],
+           'box_prompts': [[0.08, 0.18, -0.02, 0.68, 0.73, 0.315]], #, [0, 0, 0.3, 1, 1, 1.01]], #[0.11, 0.43, 0.82, 0.5, 1.01, 1.01]], 
+           'voxel_size_box': [0.04, 0.05], #0.01,
+           'mask_prompts': [[0.50455284, 0.47794762, 0.0007253083]], #[7006, 1458], , [0.28331658, 0.19435011, 0.77393067]
+           'voxel_size_mask': [0.038]
+}
+
+
+sample_1 = {'path': 'data/Objaverse/human.npy',
+           'point_prompts': [[0.57825595, 0.5005686,  0.11494722], [0.7136412,  0.49501216, 0.5020814 ], [0.7136412,  0.49501216, 0.5020814 ]], #[1112, 2133, 2133],
+           'voxel_size': [0.055, 0.045, 0.05],
+           'box_prompts': [[0., 0.17, -0.01, 0.72, 0.80, 0.3], [-0.01, 0., 0.28, 0.8, 1, 0.82], [-0.01, 0.28, 0.89, 1, 0.72, 1.02]],
+           'voxel_size_box': [0.055, 0.045, 0.055],
+           'mask_prompts': [[0.57825595, 0.5005686,  0.11494722], [0.7136412,  0.49501216, 0.5020814 ]], #[1112, 2133, 2133],
+           'voxel_size_mask': [0.055, 0.055],
+}
+sample_2 = {'path': 'data/Objaverse/lock.npy',
+           'point_prompts': [[0.6513301, 0.6753892, 0.52316076], [0.21359734, 0.6097132 , 0.7939796 ], [0.44947368, 0.21654338, 0.58450174]], #[1029, 2064, 3541],  #, [0.67447126, 0.6777649 , 0.51486933]
+           'voxel_size': [0.04, 0.05, 0.05], #, 0.05
+           'box_prompts': [[0.61, 0.4, 0.35, 0.8, 0.8, 0.6], [0.42, -0.02, -0.02, 1.02, 0.4, 1]],  #[0., 0.25, -0.02, 0.4, 0.82, 1],
+           'voxel_size_box': [0.04, 0.011],  # 0.05, 0.04
+           'mask_prompts': [[0.6513301, 0.6753892, 0.52316076], [0.21359734, 0.6097132 , 0.7939796 ], [0.9157764,  0.1995991,  0.14024617]], #[1029, 2064, 3541],
+           'voxel_size_mask': [0.04, 0.055, 0.04],
+}
+
+sample_3 = {'path': 'data/Objaverse/elephant.npy',
+           'point_prompts': [[0.4394578, 0.8342078, 0.835564]],
+           'voxel_size': [0.04],
+           'box_prompts': [[0.25,0,0,0.8,0.35,0.23]],
+           'voxel_size_box': [0.04],
+           'mask_prompts': [[0.4394578, 0.8342078, 0.835564]],
+           'voxel_size_mask': [0.04],
+}
+
+sample_4 = {'path': 'data/Objaverse/knife_rest.npy',
+           'point_prompts': [[0.3342131, 0.5378736, 0.8621972], [0.7043406, 0.4798344, 0.2585481]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.21, 0.26, 0.83, 0.37, 0.9, 1], [0, 0, 0, 1, 1, 0.28]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.3342131, 0.5378736, 0.8621972]],
+           'voxel_size_mask': [0.04],
+}
+
+sample_5 = {'path': 'data/Objaverse/skateboard.npy',
+           'point_prompts': [[0.5026503, 0.4316724, 0.5640968], [0.2835252, 0.4883442, 0.2073544]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0, 0, 0.54, 1, 1, 1], [0.21, 0.75, 0, 0.34, 1, 0.5]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.5026503, 0.4316724, 0.5640968], [0.2835252, 0.4883442, 0.2073544]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_6 = {'path': 'data/Objaverse/popcorn_machine.npy',
+           'point_prompts': [[0.278306, 0.4913014, 0.7318756], [0.5867118, 0.1180351, 0.5844101]], #, [0.8857, 0.8296, 0.6090]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.208, 0.157, 0.493, 0.779, 0.89, 0.925]],
+           'voxel_size_box': [0.04],
+           'mask_prompts': [[0.278306, 0.4913014, 0.7318756], [0.5867118, 0.1180351, 0.5844101]], #, [0.8857, 0.8296, 0.6090]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_7 = {'path': 'data/Objaverse/stove.npy',
+           'point_prompts': [[0.08, 0.72, 0.669], [0.9416, 0.3464, 0.3476], [0.021837, 0.281256, 0.8934]],
+           'voxel_size': [0.04, 0.04, 0.04],
+           'box_prompts': [[0,0,0.579,0.18,1,0.67], [0.528, 0.64, 0.508, 0.844, 0.866, 0.56]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.08, 0.72, 0.669], [0.9416, 0.3464, 0.3476], [0.021837, 0.281256, 0.8934]], 
+           'voxel_size_mask': [0.04, 0.04, 0.04],
+}
+
+
+sample_8 = {'path': 'data/Objaverse/bus_shelter.npy',
+           'point_prompts': [[0.6665938, 0.5713098, 0.2139242], [0.577489, 0.915092, 0.4498839]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.32, 0.36, 0, 0.924, 0.861, 0.394], [0, 0, 0.71, 1, 1, 1]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts': [[0.6665938, 0.5713098, 0.2139242], [0.577489, 0.915092, 0.4498839]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_9 = {'path': 'data/Objaverse/thor_hammer.npy',
+           'point_prompts': [[0.6211515, 0.5109989, 0.3867725], [0.44443, 0.2363458, 0.7229376]],
+           'voxel_size': [0.05, 0.05, 0.05],
+           'box_prompts': [[0,0,0.723,1,1,1]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.05, 0.05],
+           'mask_prompts': [[0.44443, 0.2363458, 0.7229376]],
+           'voxel_size_mask': [0.05],
+}
+
+sample_10 = {'path': 'data/Objaverse/horse.npy',
+           'point_prompts': [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.65,0,0.3,1,1,0.79], [0.37, 0, 0, 1, 1, 0.2]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts':  [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+
+sample_11 = {'path': 'data/Objaverse/dinner_booth.npy',
+           'point_prompts': [
+    [0.9192697, 0.4469184, 0.0017635],
+    [0.4987888, 0.6916906, 0.5106028]],
+           'voxel_size': [0.04, 0.04],
+           'box_prompts': [[0.65,0,0.3,1,1,0.79], [0.37, 0, 0, 1, 1, 0.2]], #, [0.353, 0.41, 0, 0.636, 0.586, 0.725]],
+           'voxel_size_box': [0.04, 0.04],
+           'mask_prompts':  [[0.3359364, 0.7555879, 0.6848574], [0.9221735, 0.1779197, 0.1927067]],
+           'voxel_size_mask': [0.04, 0.04],
+}
+# sculpture.npy
+# horse.npy
+# pipe.npy
+# dinner_booth.npy
+# ornament.npy
+# blender.npy
+# bowl.npy
+# human_face.npy
+# table.npy
+# telescope.npy
+# planet.npy
+# lamp.npy
+# dragon.npy
+
+Objaverse_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5, sample_6, sample_7, sample_8, sample_9, sample_10, sample_11]
+# sample_1, sample_2, 
+
+
+
+sample_0 = {'path': 'data/KITTI/scene1.npy',
+           'point_prompts': [[0.5527776, 0.7294311, 0.685305 ]],
+           'voxel_size': [0.02],
+           'box_prompts': [[0.52, 0.73, 0.56, 0.57, 0.76, 0.75]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5527776, 0.7294311, 0.685305 ]],
+           'voxel_size_mask': [0.02],
+}
+
+
+sample_1 = {'path': 'data/KITTI/scene2.npy',
+           'point_prompts': [[0.5090489, 0.45589063, 0.49851784]],
+           'voxel_size': [0.015],
+           'box_prompts': [[0.48, 0.43, 0.34, 0.54, 0.48, 0.71]],
+           'voxel_size_box': [0.015],
+           'mask_prompts': [[0.5090489, 0.45589063, 0.49851784]],
+           'voxel_size_mask': [0.015],
+}
+
+
+sample_2 = {'path': 'data/KITTI/scene3.npy',
+           'point_prompts': [[0.5442487, 0.5907391, 0.5992437]],
+           'voxel_size': [0.01],
+           'box_prompts': [[0.532, 0.58, 0.37, 0.555, 0.61, 0.68]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5442487, 0.5907391, 0.5992437]],
+           'voxel_size_mask': [0.01],
+}
+
+sample_3 = {'path': 'kitti/scene4.npy',
+           'point_prompts': [[0.4739189, 0.4791307, 0.8351399]],
+           'voxel_size': [0.01],
+           'box_prompts': [[0.51, 0.2, 0.75, 0.53, 0.22, 0.9]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.4739189, 0.4791307, 0.8351399], [0.4585995, 0.4209206, 0.7708794]],
+           'voxel_size_mask': [0.01, 0.006],
+}
+
+sample_4 = {'path': 'kitti/scene5.npy',
+           'point_prompts': [[0.5438917, 0.7608865, 0.5123742], [0.5131016, 0.7495122, 0.5516282]],
+           'voxel_size': [0.01, 0.01],
+           'box_prompts': [[0.43, 0.746, 0.39, 0.471,0.77, 0.62]],
+           'voxel_size_box': [0.01],
+           'mask_prompts': [[0.5438917, 0.7608865, 0.5123742], [0.5131016, 0.7495122, 0.5516282]],
+           'voxel_size_mask': [0.01, 0.01, 0.01],
+}
+
+sample_5 = {'path': 'kitti/scene6.npy',
+           'point_prompts': [[0.4619498, 0.3496694, 0.7484359], [0.4963415, 0.5221788, 0.7358279]],
+           'voxel_size': [0.008, 0.01],
+           'box_prompts': [[0.5459, 0.4, 0.62, 0.559, 0.5, 0.77], [0.61,0.343,0.625,0.664,0.377,0.8261]],
+           'voxel_size_box': [0.01, 0.01],
+           'mask_prompts': [[0.4619498, 0.3496694, 0.7484359], [0.4963415, 0.5221788, 0.7358279]],
+           'voxel_size_mask': [0.008, 0.01],
+}
+KITTI_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5]
+
+
+
+
+sample_0 = {'path': 'data/Semantic3D/scene1.npy',
+           'point_prompts': [[0.08373796, 0.61115538, 0.6007256], [0.2660193, 0.823606, 0.242315]],
+           'voxel_size': [0.017, 0.017],
+           'box_prompts': [[-0.02, 0.52, -0.02, 0.1, 0.7, 0.92]],
+           'voxel_size_box': [0.017],
+           'mask_prompts': [[0.08373796, 0.61115538, 0.6007256]],
+           'voxel_size_mask': [0.017],
+}
+
+
+sample_1 = {'path': 'data/Semantic3D/scene2.npy',
+           'point_prompts': [[0.79984724, 0.25791535, 0.18132911]],
+           'voxel_size': [0.012],
+           'box_prompts': [[0.78, 0, -0.02, 1, 0.5, 0.2]],
+           'voxel_size_box': [0.012],
+           'mask_prompts': [[0.79984724, 0.25791535, 0.18132911]],
+           'voxel_size_mask': [0.012],
+}
+
+
+
+sample_2 = {'path': 'data/Semantic3D/patch19.npy',
+           'point_prompts': [[0.51970197, 0.38389998, 0.33622117],
+                             [0.84013408, 0.80095002, 0.24210576]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [[0.51970197, 0.38389998, 0.33622117],
+                             [0.84013408, 0.80095002, 0.24210576]],
+           'voxel_size_mask': [0.017, 0.017],
+}
+
+sample_3 = {'path': 'data/Semantic3D/patch0.npy',
+           'point_prompts': [[0.91819174, 0.34150001, 0.25513778], [0., 0.34900001, 0.32881831]],
+           'voxel_size': [0.015, 0.017, 0.017, 0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+sample_4 = {'path': 'data/Semantic3D/patch1.npy',
+           'point_prompts': [[0.51603703, 0.51312565, 0.50598845]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [],
+           'voxel_size_box': [],
+           'mask_prompts': [[0.1857393, 0.2675134, 0.2463012]], #[[0.51603703, 0.51312565, 0.50598845]],
+           'voxel_size_mask': [0.01], #[0.01],
+}
+
+sample_5 = {'path': 'data/Semantic3D/patch50.npy',
+           'point_prompts': [[0.22901525, 0.49448244, 0.52076028]],
+           'voxel_size': [0.017, 0.017, 0.017, 0.017],
+           'box_prompts': [[0.09, 0.44, 0.08, 0.4, 0.75, 0.98]],
+           'voxel_size_box': [0.017, 0.017],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+
+sample_6 = {'path': 'data/Semantic3D/patch62.npy',
+           'point_prompts': [],
+           'voxel_size': [],
+           'box_prompts': [[0.26, 0.38, 0.24, 0.55, 0.78, 0.99]],
+           'voxel_size_box': [0.017],
+           'mask_prompts': [],
+           'voxel_size_mask': [],
+}
+
+Semantic3D_samples = [sample_0, sample_1, sample_2, sample_3, sample_4, sample_5, sample_6]
+
+
+
+
+VOXEL = {"point": "voxel_size", "box": "voxel_size_box", "mask": "voxel_size_mask"}
+
+

sam2point/dataset.py

@@ -0,0 +1,98 @@
+import os
+import shutil
+import numpy as np
+import scipy.io as sio
+import torch
+
+
+def load_S3DIS_sample(text_path, sample=False):
+    data = np.loadtxt(text_path)
+    point, color = data[:, :3], data[:, 3:]
+
+    point = point - point.min(axis=0)
+    point = point / point.max(axis=0)
+    color = color / 255.
+
+    return point, color
+
+def load_ScanNet_sample(data_path):
+    
+    all_data = torch.load(data_path)
+    
+    point = np.array(all_data['coord'])
+    color = np.array(all_data['color'])
+ 
+    point = point - point.min(axis=0)
+    point = point / point.max(axis=0)
+    color = color / 255.
+    return point, color
+
+def load_KITTI_sample(data_path, close=False):
+    all_data = np.load(data_path)
+    
+    point = all_data[:, :3]
+    color = all_data[:, 3:6]
+    
+    pmin = point.min(axis=0)
+    point = point - pmin
+    pmax = point.max(axis=0)
+    point = point / pmax
+    
+    # if close:
+    #     x_min, x_max = 0.3, 0.7
+    #     y_min, y_max = 0.3, 0.8
+
+    #     filter_mask = (point[:, 0] > x_min) & (point[:, 0] < x_max) & (point[:, 1] > y_min) & (point[:, 1] < y_max)
+    #     point = point[filter_mask]
+    #     color = color[filter_mask]
+
+    #     pmin = point.min(axis=0)
+    #     point = point - pmin
+    #     pmax = point.max(axis=0)
+    #     point = point / pmax
+
+    return point, color
+
+def load_Objaverse_sample(data_path):
+    all_data = np.load(data_path)
+    
+    point = all_data[:, :3]
+    color = all_data[:, 3:6]
+    
+    pmin = point.min(axis=0)
+    point = point - pmin
+    pmax = point.max(axis=0)
+    point = point / pmax
+    
+    return point, color
+
+def load_Semantic3D_sample(data_path, id, sample=False):
+    all_data = np.load(data_path)
+    
+    point = all_data[:, :3]
+    color = all_data[:, 3:6]
+    
+    pmin = point.min(axis=0)
+    point = point - pmin
+    pmax = point.max(axis=0)
+    point = point / pmax
+
+    if id > 1:  return point, color
+    if id == 0:
+        filter_mask = (point[:, 0] > 0.4) & (point[:, 1] > 0.4) & (point[:, 2] < 0.4)
+    else:
+        filter_mask = (point[:, 0] > 0.4) & (point[:, 1] < 0.5)
+    point = point[filter_mask]
+    color = color[filter_mask]
+
+    pmin = point.min(axis=0)
+    point = point - pmin
+    pmax = point.max(axis=0)
+    point = point / pmax
+
+    # if sample: #for demo
+    #     indices = np.random.choice(point.shape[0], 600000, replace=False)
+    #     point = point[indices]
+    #     color = color[indices]
+
+    return point, color

sam2point/utils.py

@@ -0,0 +1,67 @@
+import os
+import cv2
+import shutil
+import numpy as np
+import torch
+
+
+def build_fold(path):
+    if os.path.exists(path):
+        return True
+        # shutil.rmtree(path)
+        # return True 
+    os.makedirs(path)
+    return False
+    
+
+def visualize_frame_with_mask(grid0, grid1, grid2, a0_mask, a1_mask, a2_mask, point_coords, resolution, name, args=None):
+    os.makedirs(name + 'frames_seg', exist_ok=True)
+    a0_dir, a1_dir, a2_dir = name + 'frames_seg/x', name + 'frames_seg/y', name + 'frames_seg/z'
+    # build_fold(a1_dir)
+    # build_fold(a2_dir)
+    
+    a0_mask, a1_mask, a2_mask = a0_mask.repeat(1, 3, 1, 1), a1_mask.repeat(1, 3, 1, 1), a2_mask.repeat(1, 3, 1, 1)
+    a0_mask[:, 1:], a1_mask[:, 1:], a2_mask[:, 1:] = a0_mask[:, 1:] * 0, a1_mask[:, 1:] * 0, a2_mask[:, 1:] * 0
+    
+    grid0, grid1, grid2 = grid0 * 0.7 + a0_mask * 0.3, grid1 * 0.7 + a1_mask * 0.3, grid2 * 0.7 + a2_mask * 0.3
+    
+    grid0[point_coords[0], :, point_coords[1], point_coords[2]] = torch.Tensor([0., 1., 0.])
+    grid1[point_coords[1], :, point_coords[0], point_coords[2]] = torch.Tensor([0., 1., 0.])
+    grid2[point_coords[2], :, point_coords[0], point_coords[1]] = torch.Tensor([0., 1., 0.])
+    if not build_fold(a0_dir):
+        visualize_per_frame(grid0, a0_dir, resolution, args)
+    if not build_fold(a1_dir):
+        visualize_per_frame(grid1, a1_dir, resolution, args)
+    if not build_fold(a2_dir):
+        visualize_per_frame(grid2, a2_dir, resolution, args)
+    
+
+def visualize_per_frame(grid, foldpath, resolution, args=None):
+    grid = torch.nn.functional.interpolate(grid, size=(resolution, resolution), mode=args.mode)
+    # gridb = torch.nn.functional.interpolate(grid, size=(256, 256), mode='nearest')
+    # grid = grida * 0.8 + gridb * 0.2
+    
+    imgs = grid.cpu().numpy()
+    #print(imgs[0, :, 0:3, 0:3])
+    n, _, _, _ = grid.shape
+    for ii in range(n):
+        r = np.uint8(imgs[ii, 0, :, :]*255)
+        g = np.uint8(imgs[ii, 1, :, :]*255)
+        b = np.uint8(imgs[ii, 2, :, :]*255)
+        img = cv2.merge([b, g, r])
+        
+        # bilateralFilter
+        # img = cv2.bilateralFilter(img, d=-1, sigmaColor=25, sigmaSpace=7)
+        # img = cv2.bilateralFilter(img, d=9, sigmaColor=50, sigmaSpace=16)
+        # img = cv2.GaussianBlur(img, (5, 5), 0)
+        # img = cv2.medianBlur(img, 5)
+
+        cv2.imwrite('{}/{}.png'.format(foldpath, ii), img)
+    return
+
+def cal(input, points):
+    reference_point_3d = np.array(input)  
+    distances = np.linalg.norm(points - reference_point_3d, axis=1)
+    closest_index = np.argmin(distances)
+    closest_point = points[closest_index]
+    return [closest_point[0], closest_point[1], closest_point[2]]

sam2point/voxelization_utils.py

@@ -0,0 +1,150 @@
+# Please cite "4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural
+# Networks", CVPR'19 (https://arxiv.org/abs/1904.08755) if you use any part
+# of the code.
+import torch
+import numpy as np
+from collections.abc import Sequence
+
+
+def fnv_hash_vec(arr):
+    '''
+    FNV64-1A
+    '''
+    assert arr.ndim == 2
+    # Floor first for negative coordinates
+    arr = arr.copy()
+    arr = arr.astype(np.uint64, copy=False)
+    hashed_arr = np.uint64(14695981039346656037) * \
+                 np.ones(arr.shape[0], dtype=np.uint64)
+    for j in range(arr.shape[1]):
+        hashed_arr *= np.uint64(1099511628211)
+        hashed_arr = np.bitwise_xor(hashed_arr, arr[:, j])
+    return hashed_arr
+
+
+def ravel_hash_vec(arr):
+    '''
+    Ravel the coordinates after subtracting the min coordinates.
+    '''
+    assert arr.ndim == 2
+    arr = arr.copy()
+    arr -= arr.min(0)
+    arr = arr.astype(np.uint64, copy=False)
+    arr_max = arr.max(0).astype(np.uint64) + 1
+
+    keys = np.zeros(arr.shape[0], dtype=np.uint64)
+    # Fortran style indexing
+    for j in range(arr.shape[1] - 1):
+        keys += arr[:, j]
+        keys *= arr_max[j + 1]
+    keys += arr[:, -1]
+    return keys
+
+
+def sparse_quantize(coords,
+                    feats=None,
+                    labels=None,
+                    ignore_label=255,
+                    set_ignore_label_when_collision=False,
+                    return_index=False,
+                    hash_type='fnv',
+                    quantization_size=1):
+    r'''Given coordinates, and features (optionally labels), the function
+    generates quantized (voxelized) coordinates.
+
+    Args:
+        coords (:attr:`numpy.ndarray` or :attr:`torch.Tensor`): a matrix of size
+        :math:`N \times D` where :math:`N` is the number of points in the
+        :math:`D` dimensional space.
+
+        feats (:attr:`numpy.ndarray` or :attr:`torch.Tensor`, optional): a matrix of size
+        :math:`N \times D_F` where :math:`N` is the number of points and
+        :math:`D_F` is the dimension of the features.
+
+        labels (:attr:`numpy.ndarray`, optional): labels associated to eah coordinates.
+
+        ignore_label (:attr:`int`, optional): the int value of the IGNORE LABEL.
+
+        set_ignore_label_when_collision (:attr:`bool`, optional): use the `ignore_label`
+        when at least two points fall into the same cell.
+
+        return_index (:attr:`bool`, optional): True if you want the indices of the
+        quantized coordinates. False by default.
+
+        hash_type (:attr:`str`, optional): Hash function used for quantization. Either
+        `ravel` or `fnv`. `ravel` by default.
+
+        quantization_size (:attr:`float`, :attr:`list`, or
+        :attr:`numpy.ndarray`, optional): the length of the each side of the
+        hyperrectangle of of the grid cell.
+
+    .. note::
+        Please check `examples/indoor.py` for the usage.
+
+    '''
+    use_label = labels is not None
+    use_feat = feats is not None
+    if not use_label and not use_feat:
+        return_index = True
+
+    assert hash_type in [
+        'ravel', 'fnv'
+    ], "Invalid hash_type. Either ravel, or fnv allowed. You put hash_type=" + hash_type
+    assert coords.ndim == 2, \
+        "The coordinates must be a 2D matrix. The shape of the input is " + str(coords.shape)
+    if use_feat:
+        assert feats.ndim == 2
+        assert coords.shape[0] == feats.shape[0]
+    if use_label:
+        assert coords.shape[0] == len(labels)
+
+    # Quantize the coordinates
+    dimension = coords.shape[1]
+    if isinstance(quantization_size, (Sequence, np.ndarray, torch.Tensor)):
+        assert len(
+            quantization_size
+        ) == dimension, "Quantization size and coordinates size mismatch."
+        quantization_size = [i for i in quantization_size]
+    elif np.isscalar(quantization_size):  # Assume that it is a scalar
+        quantization_size = [quantization_size for i in range(dimension)]
+    else:
+        raise ValueError('Not supported type for quantization_size.')
+    discrete_coords = np.floor(coords / np.array(quantization_size))
+
+    # Hash function type
+    if hash_type == 'ravel':
+        key = ravel_hash_vec(discrete_coords)
+    else:
+        key = fnv_hash_vec(discrete_coords)
+
+    if use_label:
+        _, inds, counts = np.unique(key, return_index=True, return_counts=True)
+        filtered_labels = labels[inds]
+        if set_ignore_label_when_collision:
+            filtered_labels[counts > 1] = ignore_label
+        if return_index:
+            return inds, filtered_labels
+        else:
+            return discrete_coords[inds], feats[inds], filtered_labels
+    else:
+        _, inds, inds_reverse = np.unique(key, return_index=True, return_inverse=True)
+        # NOTE:
+        if use_feat:
+            voxel_feats = np.zeros((len(np.unique(key)), feats.shape[1]), dtype=feats.dtype)
+            for i in range(len(np.unique(key))):
+            #     voxel_feats[i] = np.mean(feats[inds_reverse == i], axis=0)
+            #     voxel_feats[i] = np.median(feats[inds_reverse == i], axis=0)
+                voxel_center = np.mean(coords[inds_reverse == i], axis=0)
+                distances = np.linalg.norm(coords[inds_reverse == i] - voxel_center, axis=1)
+                central_point_idx = np.argmin(distances)
+                voxel_feats[i] = feats[inds_reverse == i][central_point_idx]
+            if return_index:
+                return inds, inds_reverse, voxel_feats
+        ##############
+        if return_index:
+            return inds, inds_reverse
+        else:
+            if use_feat:
+                return discrete_coords[inds], feats[inds]
+            else:
+                return discrete_coords[inds]

sam2point/voxelizer.py

@@ -0,0 +1,89 @@
+# Codes are taken from BPNet, CVPR'21
+# https://github.com/wbhu/BPNet/blob/main/dataset/voxelizer.py
+
+import collections
+import numpy as np
+from sam2point.voxelization_utils import sparse_quantize
+from scipy.linalg import expm, norm
+
+
+# Rotation matrix along axis with angle theta
+def M(axis, theta):
+    return expm(np.cross(np.eye(3), axis / norm(axis) * theta))
+
+
+class Voxelizer:
+
+    def __init__(self, voxel_size=1, clip_bound=None):
+        '''
+        Args:
+          voxel_size: side length of a voxel
+          clip_bound: boundary of the voxelizer. Points outside the bound will be deleted
+            expects either None or an array like ((-100, 100), (-100, 100), (-100, 100)).
+          ignore_label: label assigned for ignore (not a training label).
+        '''
+        self.voxel_size = voxel_size
+        self.clip_bound = clip_bound
+
+    def get_transformation_matrix(self):
+        voxelization_matrix = np.eye(4)
+
+        # Transform pointcloud coordinate to voxel coordinate.
+        scale = 1 / self.voxel_size
+        np.fill_diagonal(voxelization_matrix[:3, :3], scale)
+        # Get final transformation matrix.
+        return voxelization_matrix
+
+    def clip(self, coords, center=None):
+        bound_min = np.min(coords, 0).astype(float)
+        bound_max = np.max(coords, 0).astype(float)
+        bound_size = bound_max - bound_min
+        if center is None:
+            center = bound_min + bound_size * 0.5
+        lim = self.clip_bound
+        
+        # Clip points outside the limit
+        clip_inds = ((coords[:, 0] >= (lim[0][0] + center[0])) &
+                     (coords[:, 0] < (lim[0][1] + center[0])) &
+                     (coords[:, 1] >= (lim[1][0] + center[1])) &
+                     (coords[:, 1] < (lim[1][1] + center[1])) &
+                     (coords[:, 2] >= (lim[2][0] + center[2])) &
+                     (coords[:, 2] < (lim[2][1] + center[2])))
+        return clip_inds
+
+    def voxelize(self, coords, feats, labels, center=None, link=None, return_ind=False):
+        assert coords.shape[1] == 3 and coords.shape[0] == feats.shape[0] and coords.shape[0]
+        if self.clip_bound is not None:
+            clip_inds = self.clip(coords, center)
+            if clip_inds.sum():
+                coords, feats = coords[clip_inds], feats[clip_inds]
+                if labels is not None:
+                    labels = labels[clip_inds]
+
+        # Get rotation and scale
+        M_v = self.get_transformation_matrix()
+        # Apply transformations
+        rigid_transformation = M_v
+        homo_coords = np.hstack((coords, np.ones((coords.shape[0], 1), dtype=coords.dtype)))
+        coords_aug = np.floor(homo_coords @ rigid_transformation.T[:, :3])
+
+        # Align all coordinates to the origin.
+        min_coords = coords_aug.min(0)
+        M_t = np.eye(4)
+        M_t[:3, -1] = -min_coords
+        rigid_transformation = M_t @ rigid_transformation
+        coords_aug = np.floor(coords_aug - min_coords)
+
+        inds, inds_reconstruct = sparse_quantize(coords_aug, return_index=True) #NOTE
+        coords_aug, feats, labels = coords_aug[inds], feats[inds], labels[inds] #NOTE
+
+        # #NOTE:
+        # inds, inds_reconstruct, feats = sparse_quantize(coords_aug, feats=feats, return_index=True)
+        # coords_aug, labels = coords_aug[inds], labels[inds]
+
+        if return_ind:
+            return coords_aug, feats, labels, np.array(inds_reconstruct), inds
+        if link is not None:
+            return coords_aug, feats, labels, np.array(inds_reconstruct), link[inds]
+
+        return coords_aug, feats, labels, np.array(inds_reconstruct)

sam2point/voxelizer0.py

@@ -0,0 +1,140 @@
+# Codes are taken from BPNet, CVPR'21
+# https://github.com/wbhu/BPNet/blob/main/dataset/voxelizer.py
+
+import collections
+import numpy as np
+from sam2point.voxelization_utils import sparse_quantize
+from scipy.linalg import expm, norm
+
+
+# Rotation matrix along axis with angle theta
+def M(axis, theta):
+    return expm(np.cross(np.eye(3), axis / norm(axis) * theta))
+
+
+class Voxelizer:
+
+    def __init__(self,
+                 voxel_size=1,
+                 clip_bound=None,
+                 use_augmentation=False,
+                 scale_augmentation_bound=None,
+                 rotation_augmentation_bound=None,
+                 translation_augmentation_ratio_bound=None,
+                 ignore_label=255):
+        '''
+        Args:
+          voxel_size: side length of a voxel
+          clip_bound: boundary of the voxelizer. Points outside the bound will be deleted
+            expects either None or an array like ((-100, 100), (-100, 100), (-100, 100)).
+          scale_augmentation_bound: None or (0.9, 1.1)
+          rotation_augmentation_bound: None or ((np.pi / 6, np.pi / 6), None, None) for 3 axis.
+            Use random order of x, y, z to prevent bias.
+          translation_augmentation_bound: ((-5, 5), (0, 0), (-10, 10))
+          ignore_label: label assigned for ignore (not a training label).
+        '''
+        self.voxel_size = voxel_size
+        self.clip_bound = clip_bound
+        self.ignore_label = ignore_label
+
+        # Augmentation
+        self.use_augmentation = use_augmentation
+        self.scale_augmentation_bound = scale_augmentation_bound
+        self.rotation_augmentation_bound = rotation_augmentation_bound
+        self.translation_augmentation_ratio_bound = translation_augmentation_ratio_bound
+
+    def get_transformation_matrix(self):
+        voxelization_matrix, rotation_matrix = np.eye(4), np.eye(4)
+        # Get clip boundary from config or pointcloud.
+        # Get inner clip bound to crop from.
+
+        # Transform pointcloud coordinate to voxel coordinate.
+        # 1. Random rotation
+        rot_mat = np.eye(3)
+        if self.use_augmentation and self.rotation_augmentation_bound is not None:
+            if isinstance(self.rotation_augmentation_bound, collections.Iterable):
+                rot_mats = []
+                for axis_ind, rot_bound in enumerate(self.rotation_augmentation_bound):
+                    theta = 0
+                    axis = np.zeros(3)
+                    axis[axis_ind] = 1
+                    if rot_bound is not None:
+                        theta = np.random.uniform(*rot_bound)
+                    rot_mats.append(M(axis, theta))
+                # Use random order
+                np.random.shuffle(rot_mats)
+                rot_mat = rot_mats[0] @ rot_mats[1] @ rot_mats[2]
+            else:
+                raise ValueError()
+        rotation_matrix[:3, :3] = rot_mat
+        # 2. Scale and translate to the voxel space.
+        scale = 1 / self.voxel_size
+        if self.use_augmentation and self.scale_augmentation_bound is not None:
+            scale *= np.random.uniform(*self.scale_augmentation_bound)
+        np.fill_diagonal(voxelization_matrix[:3, :3], scale)
+        # Get final transformation matrix.
+        return voxelization_matrix, rotation_matrix
+
+    def clip(self, coords, center=None, trans_aug_ratio=None):
+        bound_min = np.min(coords, 0).astype(float)
+        bound_max = np.max(coords, 0).astype(float)
+        bound_size = bound_max - bound_min
+        if center is None:
+            center = bound_min + bound_size * 0.5
+        lim = self.clip_bound
+        if trans_aug_ratio is not None:
+            trans = np.multiply(trans_aug_ratio, bound_size)
+            center += trans
+        # Clip points outside the limit
+        clip_inds = ((coords[:, 0] >= (lim[0][0] + center[0])) &
+                     (coords[:, 0] < (lim[0][1] + center[0])) &
+                     (coords[:, 1] >= (lim[1][0] + center[1])) &
+                     (coords[:, 1] < (lim[1][1] + center[1])) &
+                     (coords[:, 2] >= (lim[2][0] + center[2])) &
+                     (coords[:, 2] < (lim[2][1] + center[2])))
+        return clip_inds
+
+    def voxelize(self, coords, feats, labels, center=None, link=None, return_ind=False):
+        assert coords.shape[1] == 3 and coords.shape[0] == feats.shape[0] and coords.shape[0]
+        if self.clip_bound is not None:
+            trans_aug_ratio = np.zeros(3)
+            if self.use_augmentation and self.translation_augmentation_ratio_bound is not None:
+                for axis_ind, trans_ratio_bound in enumerate(self.translation_augmentation_ratio_bound):
+                    trans_aug_ratio[axis_ind] = np.random.uniform(*trans_ratio_bound)
+
+            clip_inds = self.clip(coords, center, trans_aug_ratio)
+            if clip_inds.sum():
+                coords, feats = coords[clip_inds], feats[clip_inds]
+                if labels is not None:
+                    labels = labels[clip_inds]
+
+        # Get rotation and scale
+        M_v, M_r = self.get_transformation_matrix()
+        # Apply transformations
+        rigid_transformation = M_v
+        if self.use_augmentation:
+            rigid_transformation = M_r @ rigid_transformation
+
+        homo_coords = np.hstack((coords, np.ones((coords.shape[0], 1), dtype=coords.dtype)))
+        coords_aug = np.floor(homo_coords @ rigid_transformation.T[:, :3])
+
+        # Align all coordinates to the origin.
+        min_coords = coords_aug.min(0)
+        M_t = np.eye(4)
+        M_t[:3, -1] = -min_coords
+        rigid_transformation = M_t @ rigid_transformation
+        coords_aug = np.floor(coords_aug - min_coords)
+
+        inds, inds_reconstruct = sparse_quantize(coords_aug, return_index=True)
+        coords_aug, feats, labels = coords_aug[inds], feats[inds], labels[inds]
+
+        # Normal rotation
+        if feats.shape[1] > 6:
+            feats[:, 3:6] = feats[:, 3:6] @ (M_r[:3, :3].T)
+
+        if return_ind:
+            return coords_aug, feats, labels, np.array(inds_reconstruct), inds
+        if link is not None:
+            return coords_aug, feats, labels, np.array(inds_reconstruct), link[inds]
+
+        return coords_aug, feats, labels, np.array(inds_reconstruct)

segment.py

@@ -0,0 +1,351 @@
+from PIL import Image
+import numpy as np
+import torch, os
+import sam2point.utils as utils
+from sam2.build_sam import build_sam2_video_predictor, build_sam2
+from sam2.sam2_image_predictor import SAM2ImagePredictor
+
+CHECKPOINT = "./checkpoints/sam2_hiera_large.pt"
+MODELCFG = "sam2_hiera_l.yaml"
+RESOLUTION = 256
+
+def grid_to_frames(grid, foldpath, args):
+    if not utils.build_fold(foldpath):
+        utils.visualize_per_frame(grid, foldpath=foldpath, resolution=RESOLUTION, args=args)
+    
+    # scan all the JPEG frame names in this directory
+    frame_names = [
+        p for p in os.listdir(foldpath)
+        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG", ".png"]
+    ]
+    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    
+    for i in range(len(frame_names)):
+        frame_names[i] = os.path.join(foldpath, frame_names[i])
+    
+    return frame_names
+
+
+def segment_point(frame_paths, point):
+    sam2_checkpoint = CHECKPOINT
+    model_cfg = MODELCFG
+    
+    predictor = build_sam2_video_predictor(model_cfg, sam2_checkpoint)
+    inference_state = predictor.init_state(frame_paths=frame_paths)
+    
+    predictor.reset_state(inference_state)
+    
+    ann_frame_idx = 0  # the frame index we interact with
+    ann_obj_id = 1  # give a unique id to each object we interact with (it can be any integers)
+
+    # Let's add a positive click at (x, y) = (210, 350) to get started
+    points = np.array([point], dtype=np.float32)
+    # for labels, `1` means positive click and `0` means negative click
+    labels = np.array([1], np.int32)
+    _, out_obj_ids, out_mask_logits = predictor.add_new_points_or_box(
+        inference_state=inference_state,
+        frame_idx=ann_frame_idx,
+        obj_id=ann_obj_id,
+        points=points,
+        labels=labels,
+    )
+    
+    # run propagation throughout the video and collect the results in a dict
+    video_segments = {}  # video_segments contains the per-frame segmentation results
+    for out_frame_idx, out_obj_ids, out_mask_logits in predictor.propagate_in_video(inference_state):
+        video_segments[out_frame_idx] = {
+            out_obj_id: (out_mask_logits[i] > 0.0).cpu().numpy()
+            for i, out_obj_id in enumerate(out_obj_ids)
+        }
+    
+    masks = []
+    for out_frame_idx in range(0, len(frame_paths)):       
+        for out_obj_id, out_mask in video_segments[out_frame_idx].items():
+            out_mask = torch.from_numpy(out_mask * 1.0)
+            masks.append(out_mask)
+    masks = torch.cat(masks, dim=0)
+    return masks
+            
+
+
+def segment_box(frame_paths, box, n_frame):
+    sam2_checkpoint = CHECKPOINT
+    model_cfg = MODELCFG
+        
+    predictor = build_sam2_video_predictor(model_cfg, sam2_checkpoint)
+    inference_state = predictor.init_state(frame_paths=frame_paths)
+    
+    predictor.reset_state(inference_state)
+    
+    for i in range(n_frame):
+    
+        ann_frame_idx = i  # the frame index we interact with
+        ann_obj_id = 1  # give a unique id to each object we interact with (it can be any integers)
+
+        # Let's add a positive click at (x, y) = (210, 350) to get started
+        box = np.array(box, dtype=np.float32)
+        # for labels, `1` means positive click and `0` means negative click
+        _, out_obj_ids, out_mask_logits = predictor.add_new_points_or_box(
+            inference_state=inference_state,
+            frame_idx=ann_frame_idx,
+            obj_id=ann_obj_id,
+            box=box,
+        )
+    
+    # run propagation throughout the video and collect the results in a dict
+    video_segments = {}  # video_segments contains the per-frame segmentation results
+    for out_frame_idx, out_obj_ids, out_mask_logits in predictor.propagate_in_video(inference_state):
+        video_segments[out_frame_idx] = {
+            out_obj_id: (out_mask_logits[i] > 0.0).cpu().numpy()
+            for i, out_obj_id in enumerate(out_obj_ids)
+        }
+    
+    masks = []
+    for out_frame_idx in range(0, len(frame_paths)):       
+        for out_obj_id, out_mask in video_segments[out_frame_idx].items():
+            out_mask = torch.from_numpy(out_mask * 1.0)
+            masks.append(out_mask)
+    masks = torch.cat(masks, dim=0)
+    # print(masks.shape)
+    return masks
+            
+
+def segment_mask(frame_paths, point):
+    
+    sam2_checkpoint = CHECKPOINT
+    model_cfg = MODELCFG
+    
+    # generate a mask for one frame, where we use the image predictor
+    sam2_image_model = build_sam2(model_cfg, sam2_checkpoint)
+    image_predictor = SAM2ImagePredictor(sam2_image_model)
+    image = Image.open(frame_paths[0])
+    image_predictor.set_image(np.array(image.convert("RGB")))
+    
+    point = np.array([point], dtype=np.float32)
+    label = np.array([1], np.int32)
+    masks, scores, logits = image_predictor.predict(point_coords=point, point_labels=label, multimask_output=True)
+    sorted_ind = np.argsort(scores)[::-1]
+    masks = masks[sorted_ind]
+    
+    
+    # predict the mask for other frames
+    video_predictor = build_sam2_video_predictor(model_cfg, sam2_checkpoint)
+    inference_state = video_predictor.init_state(frame_paths=frame_paths)
+    
+    video_predictor.reset_state(inference_state)
+    
+    ann_frame_idx = 0  # the frame index we interact with
+    ann_obj_id = 1  # give a unique id to each object we interact with (it can be any integers)
+    
+    mask_prompt = masks[0]
+    video_predictor.add_new_mask(inference_state=inference_state, frame_idx=ann_frame_idx, obj_id=ann_obj_id, mask=mask_prompt)
+    # run propagation throughout the video and collect the results in a dict
+    video_segments = {}  # video_segments contains the per-frame segmentation results
+    for out_frame_idx, out_obj_ids, out_mask_logits in video_predictor.propagate_in_video(inference_state):
+        video_segments[out_frame_idx] = {
+            out_obj_id: (out_mask_logits[i] > 0.0).cpu().numpy()
+            for i, out_obj_id in enumerate(out_obj_ids)
+        }
+    
+    masks = []
+    for out_frame_idx in range(0, len(frame_paths)):       
+        for out_obj_id, out_mask in video_segments[out_frame_idx].items():
+            out_mask = torch.from_numpy(out_mask * 1.0)
+            masks.append(out_mask)
+    masks = torch.cat(masks, dim=0)
+    
+    return masks, mask_prompt
+                            
+def seg_point(locs, feats, prompt, args):
+    num_voxels = locs.max().astype(int)
+    grid = np.ones((num_voxels + 5, num_voxels+5, num_voxels+5, 3))
+    
+    # padding 
+    locs = locs.astype(int)
+    for v in range(locs.shape[0]):
+        grid[locs[v][0]+2,locs[v][1]+2,locs[v][2]+2] = feats[v]
+    
+    X, Y, Z, _ = grid.shape
+    grid = torch.from_numpy(grid)
+    
+    name_list = ["./tmp/" + args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+    name = '_'.join(name_list)
+    os.makedirs(name + 'frames', exist_ok=True)
+    axis0, axis1, axis2 = name + "frames/x", name + "frames/y", name + "frames/z"
+    grid0, grid1, grid2 = grid.permute(0,3,1,2), grid.permute(1,3,0,2), grid.permute(2,3,0,1)
+
+    a0_frame_paths = grid_to_frames(grid0, axis0, args)
+    a1_frame_paths = grid_to_frames(grid1, axis1, args)
+    a2_frame_paths = grid_to_frames(grid2, axis2, args)
+    
+    voxel_coords = np.array(prompt) / args.voxel_size + 2
+    voxel_coords = voxel_coords.astype(int) 
+
+    pixel = voxel_coords * 1.0 / X * RESOLUTION + args.theta * RESOLUTION / X
+    pixel = pixel.astype(int)
+
+    idx = args.prompt_idx
+    a0_paths_0, a0_paths_1 = a0_frame_paths[:voxel_coords[idx, 0]+1][::-1], a0_frame_paths[voxel_coords[idx, 0]:]
+    a1_paths_0, a1_paths_1 = a1_frame_paths[:voxel_coords[idx, 1]+1][::-1], a1_frame_paths[voxel_coords[idx, 1]:]
+    a2_paths_0, a2_paths_1 = a2_frame_paths[:voxel_coords[idx, 2]+1][::-1], a2_frame_paths[voxel_coords[idx, 2]:]
+    
+    a0_mask_0 = torch.flip(segment_point(a0_paths_0, [pixel[idx, 2], pixel[idx, 1]]), dims=[0])
+    a0_mask_1 = segment_point(a0_paths_1, [pixel[idx, 2], pixel[idx, 1]])[1:, :, :]
+    a0_mask = torch.cat([a0_mask_0, a0_mask_1], dim=0)
+    a0_mask = torch.nn.functional.interpolate(a0_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a1_mask_0 = torch.flip(segment_point(a1_paths_0, [pixel[idx, 2], pixel[idx, 0]]), dims=[0])
+    a1_mask_1 = segment_point(a1_paths_1, [pixel[idx, 2], pixel[idx, 0]])[1:, :, :]
+    a1_mask = torch.cat([a1_mask_0, a1_mask_1], dim=0)
+    a1_mask = torch.nn.functional.interpolate(a1_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a2_mask_0 = torch.flip(segment_point(a2_paths_0, [pixel[idx, 1], pixel[idx, 0]]), dims=[0])
+    a2_mask_1 = segment_point(a2_paths_1, [pixel[idx, 1], pixel[idx, 0]])[1:, :, :]
+    a2_mask = torch.cat([a2_mask_0, a2_mask_1], dim=0)
+    a2_mask = torch.nn.functional.interpolate(a2_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a0_mask, a1_mask, a2_mask = a0_mask.transpose(0, 1), a1_mask.transpose(0, 1), a2_mask.transpose(0, 1)
+    # utils.visualize_frame_with_mask(grid0, grid1, grid2, a0_mask, a1_mask, a2_mask, voxel_coords[idx], resolution=RESOLUTION)
+    
+    mask = a0_mask.permute(0, 2, 3, 1) + a1_mask.permute(2, 0, 3, 1) + a2_mask.permute(2, 3, 0, 1)
+    mask = (mask > 1.5).squeeze()[2:, 2:, 2:]
+    return mask
+
+def seg_box(locs, feats, prompt, args):
+    num_voxels = locs.max().astype(int)
+    grid = np.ones((num_voxels + 5, num_voxels+5, num_voxels+5, 3))
+    
+    # padding 
+    locs = locs.astype(int)
+    for v in range(locs.shape[0]):
+        grid[locs[v][0]+2,locs[v][1]+2,locs[v][2]+2] = feats[v]
+    
+    X, Y, Z, _ = grid.shape
+    grid = torch.from_numpy(grid)
+
+    name_list = ["./tmp/" + args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+    name = '_'.join(name_list)     
+    os.makedirs(name + 'frames', exist_ok=True)
+    axis0, axis1, axis2 = name + "frames/x", name + "frames/y", name + "frames/z"
+    grid0, grid1, grid2 = grid.permute(0,3,1,2), grid.permute(1,3,0,2), grid.permute(2,3,0,1)
+
+    a0_frame_paths = grid_to_frames(grid0, axis0, args)
+    a1_frame_paths = grid_to_frames(grid1, axis1, args)
+    a2_frame_paths = grid_to_frames(grid2, axis2, args)
+    
+    point_prompts = np.array(prompt)
+    voxel_coords = point_prompts / args.voxel_size + 2
+    voxel_coords = voxel_coords.astype(int)
+    
+    pixel = voxel_coords * 1.0 / X * RESOLUTION + args.theta * RESOLUTION / X
+    pixel = pixel.astype(int)
+        
+    idx = args.prompt_idx
+    a0_paths_0, a0_paths_1 = a0_frame_paths[:voxel_coords[idx, 3]+1][::-1], a0_frame_paths[voxel_coords[idx, 0]:]
+    a1_paths_0, a1_paths_1 = a1_frame_paths[:voxel_coords[idx, 4]+1][::-1], a1_frame_paths[voxel_coords[idx, 1]:]
+    a2_paths_0, a2_paths_1 = a2_frame_paths[:voxel_coords[idx, 5]+1][::-1], a2_frame_paths[voxel_coords[idx, 2]:]
+    
+    frame_num0 = voxel_coords[idx, 3] - voxel_coords[idx, 0]
+    end0, start0 = len(a0_paths_0) - int(frame_num0 / 2), int(frame_num0 / 2)
+    a0_mask_0 = torch.flip(segment_box(a0_paths_0, [pixel[idx, 2], pixel[idx, 1], pixel[idx, 5], pixel[idx, 4]], frame_num0), dims=[0])[:end0]
+    a0_mask_1 = segment_box(a0_paths_1, [pixel[idx, 2], pixel[idx, 1], pixel[idx, 5], pixel[idx, 4]], frame_num0)[start0:]
+    a0_mask = torch.cat([a0_mask_0, a0_mask_1], dim=0)
+    a0_mask = torch.nn.functional.interpolate(a0_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    frame_num1 = voxel_coords[idx, 4] - voxel_coords[idx, 1]
+    end1, start1 = len(a1_paths_0) - int(frame_num1 / 2), int(frame_num1 / 2)
+    a1_mask_0 = torch.flip(segment_box(a1_paths_0, [pixel[idx, 2], pixel[idx, 0], pixel[idx, 5], pixel[idx, 3]], frame_num1), dims=[0])[:end1]
+    a1_mask_1 = segment_box(a1_paths_1, [pixel[idx, 2], pixel[idx, 0], pixel[idx, 5], pixel[idx, 3]], frame_num1)[start1:]
+    a1_mask = torch.cat([a1_mask_0, a1_mask_1], dim=0)
+    a1_mask = torch.nn.functional.interpolate(a1_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    frame_num2 = voxel_coords[idx, 5] - voxel_coords[idx, 2]
+    end2, start2 = len(a2_paths_0) - int(frame_num2 / 2), int(frame_num2 / 2)
+    a2_mask_0 = torch.flip(segment_box(a2_paths_0, [pixel[idx, 1], pixel[idx, 0], pixel[idx, 4], pixel[idx, 3]], frame_num2), dims=[0])[:end2]
+    a2_mask_1 = segment_box(a2_paths_1, [pixel[idx, 1], pixel[idx, 0], pixel[idx, 4], pixel[idx, 3]], frame_num2)[start2:]
+    a2_mask = torch.cat([a2_mask_0, a2_mask_1], dim=0)
+    a2_mask = torch.nn.functional.interpolate(a2_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a0_mask, a1_mask, a2_mask = a0_mask.transpose(0, 1), a1_mask.transpose(0, 1), a2_mask.transpose(0, 1)
+    # utils.visualize_frame_with_mask(grid0, grid1, grid2, a0_mask, a1_mask, a2_mask, voxel_coords[idx], resolution=RESOLUTION)
+    
+    mask = a0_mask.permute(0, 2, 3, 1) + a1_mask.permute(2, 0, 3, 1) + a2_mask.permute(2, 3, 0, 1)
+    mask = (mask > 1.5).squeeze()[2:, 2:, 2:]
+    
+    return mask
+
+def seg_mask(locs, feats, prompt, args):
+    num_voxels = locs.max().astype(int)
+    grid = np.ones((num_voxels + 5, num_voxels+5, num_voxels+5, 3))
+    
+    # padding 
+    locs = locs.astype(int)
+    for v in range(locs.shape[0]):
+        grid[locs[v][0]+2,locs[v][1]+2,locs[v][2]+2] = feats[v]
+    
+    X, Y, Z, _ = grid.shape
+    grid = torch.from_numpy(grid)
+
+    name_list = ["./tmp/" + args.dataset, "sample" + str(args.sample_idx), args.prompt_type + "-prompt" + str(args.prompt_idx)]
+    name = '_'.join(name_list)
+    os.makedirs(name + 'frames', exist_ok=True)
+    axis0, axis1, axis2 = name + "frames/x", name + "frames/y", name + "frames/z"
+    grid0, grid1, grid2 = grid.permute(0,3,1,2), grid.permute(1,3,0,2), grid.permute(2,3,0,1)
+
+    a0_frame_paths = grid_to_frames(grid0, axis0, args)
+    a1_frame_paths = grid_to_frames(grid1, axis1, args)
+    a2_frame_paths = grid_to_frames(grid2, axis2, args)
+    
+    point_prompts = np.array(prompt)
+    voxel_coords = point_prompts / args.voxel_size + 2
+    voxel_coords = voxel_coords.astype(int)
+    
+    pixel = voxel_coords * 1.0 / X * RESOLUTION + args.theta * RESOLUTION / X
+    pixel = pixel.astype(int)
+        
+    idx = args.prompt_idx
+    a0_paths_0, a0_paths_1 = a0_frame_paths[:voxel_coords[idx, 0]+1][::-1], a0_frame_paths[voxel_coords[idx, 0]:]
+    a1_paths_0, a1_paths_1 = a1_frame_paths[:voxel_coords[idx, 1]+1][::-1], a1_frame_paths[voxel_coords[idx, 1]:]
+    a2_paths_0, a2_paths_1 = a2_frame_paths[:voxel_coords[idx, 2]+1][::-1], a2_frame_paths[voxel_coords[idx, 2]:]
+    
+    a0_mask_0, a0_prompt = segment_mask(a0_paths_0, [pixel[idx, 2], pixel[idx, 1]])
+    a0_mask_0 = torch.flip(a0_mask_0, dims=[0])
+    a0_mask_1, _ = segment_mask(a0_paths_1, [pixel[idx, 2], pixel[idx, 1]])
+    a0_mask_1 = a0_mask_1[1:, :, :]
+    a0_mask = torch.cat([a0_mask_0, a0_mask_1], dim=0)
+    a0_prompt_mask = a0_mask * 0
+    a0_prompt_mask[voxel_coords[idx, 0]] = torch.from_numpy(a0_prompt)
+    a0_mask = torch.nn.functional.interpolate(a0_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    a0_prompt_mask = torch.nn.functional.interpolate(a0_prompt_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a1_mask_0, a1_prompt = segment_mask(a1_paths_0, [pixel[idx, 2], pixel[idx, 0]])
+    a1_mask_0 = torch.flip(a1_mask_0, dims=[0])
+    a1_mask_1, _ = segment_mask(a1_paths_1, [pixel[idx, 2], pixel[idx, 0]])
+    a1_mask_1 = a1_mask_1[1:, :, :]
+    a1_mask = torch.cat([a1_mask_0, a1_mask_1], dim=0)
+    a1_prompt_mask = a1_mask * 0
+    a1_prompt_mask[voxel_coords[idx, 1]] = torch.from_numpy(a1_prompt)
+    a1_mask = torch.nn.functional.interpolate(a1_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    a1_prompt_mask = torch.nn.functional.interpolate(a1_prompt_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a2_mask_0, a2_prompt = segment_mask(a2_paths_0, [pixel[idx, 1], pixel[idx, 0]])
+    a2_mask_0 = torch.flip(a2_mask_0, dims=[0])
+    a2_mask_1, _ = segment_mask(a2_paths_1, [pixel[idx, 1], pixel[idx, 0]])
+    a2_mask_1 = a2_mask_1[1:, :, :]
+    a2_mask = torch.cat([a2_mask_0, a2_mask_1], dim=0)
+    a2_prompt_mask = a2_mask * 0
+    a2_prompt_mask[voxel_coords[idx, 2]] = torch.from_numpy(a2_prompt)
+    a2_mask = torch.nn.functional.interpolate(a2_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    a2_prompt_mask = torch.nn.functional.interpolate(a2_prompt_mask.unsqueeze(0).unsqueeze(0), size=(X, X, X), mode='trilinear').squeeze(0)
+    
+    a0_mask, a1_mask, a2_mask = a0_mask.transpose(0, 1), a1_mask.transpose(0, 1), a2_mask.transpose(0, 1)
+    utils.visualize_frame_with_mask(grid0, grid1, grid2, a0_mask, a1_mask, a2_mask, voxel_coords[idx], resolution=RESOLUTION, name=name, args=args)
+    a0_prompt_mask, a1_prompt_mask, a2_prompt_mask = a0_prompt_mask.transpose(0, 1), a1_prompt_mask.transpose(0, 1), a2_prompt_mask.transpose(0, 1)
+    
+    mask = a0_mask.permute(0, 2, 3, 1) + a1_mask.permute(2, 0, 3, 1) + a2_mask.permute(2, 3, 0, 1)
+    mask = (mask > 1.5).squeeze()[2:, 2:, 2:]
+    
+    prompt_mask = a0_prompt_mask.permute(0, 2, 3, 1) + a1_prompt_mask.permute(2, 0, 3, 1) + a2_prompt_mask.permute(2, 3, 0, 1)
+    prompt_mask = (prompt_mask > 0.5).squeeze()[2:, 2:, 2:]
+    
+    return mask, prompt_mask