Initial upload

README.md

@@ -0,0 +1,135 @@
+<p align="center">
+<h1 align="center">Map It Anywhere (MIA): Empowering Bird’s Eye View Mapping using Large-scale Public Data</h1>
+
+  <p align="center">
+    <a href="https://cherieho.com/"><strong>Cherie Ho*</strong></a>
+    ·
+    <a href="https://www.linkedin.com/in/tonyjzou/"><strong>Jiaye (Tony) Zou*</strong></a>
+    ·
+    <a href="https://www.linkedin.com/in/omaralama/"><strong>Omar Alama*</strong></a>
+    <br>
+    <a href="https://smj007.github.io/"><strong>Sai Mitheran Jagadesh Kumar</strong></a>
+    ·
+    <a href="https://github.com/chychiang"><strong>Benjamin Chiang</strong></a>
+    ·
+    <a href="https://www.linkedin.com/in/taneesh-gupta/"><strong>Taneesh Gupta</strong></a>
+    ·
+    <a href="https://sairlab.org/team/chenw/"><strong>Chen Wang</strong></a>
+    <br>
+    <a href="https://nik-v9.github.io/"><strong>Nikhil Keetha</strong></a>
+    ·
+    <a href="https://www.cs.cmu.edu/~./katia/"><strong>Katia Sycara</strong></a>
+    ·
+    <a href="https://theairlab.org/team/sebastian/"><strong>Sebastian Scherer</strong></a>
+    <br>
+  </p>
+
+</p>
+
+![Map It Anywhere (MIA)](/assets/mia_pull_fig.png "Map It Anywhere (MIA)")
+
+## Table of Contents
+  - [Using the MIA Data Engine](#using-the-mia-data-engine)
+  - [Downloading the MIA dataset](#downloading-the-mia-dataset)
+  - [Training](#training)
+  - [Evaluation](#evaluation)
+  - [Acknowledgement](#acknowledgement)
+
+
+## Using the MIA data engine
+
+### 0. Setting up the environment
+0. Install docker by following the instructions on their [website](https://www.docker.com/get-started/)
+1. Build the docker image `mia/Dockerfile` by running: 
+
+        docker build -t mia:release mia/Dockerfile
+2. Launch the container while mounting this repository to the container file system.
+
+        docker run -v <PATH_TO_THIS_REPO>:/home/MapItAnywhere --network=bridge -it mia:release
+
+### 1. Getting FPVs
+
+The first stage of the MIA data engine is to get the first person images.
+First, if you want to pull your own locations, copy the example configuration from `mia/conf/example.yaml` and edit the cities list to specify the cities you want. Feel free to explore the other well-documented FPV options in the configuration file.
+
+Once configuration is done simply run the following from inside your docker container with working dir set to this repo:
+
+    python3.9 -m mia.fpv.get_fpv --cfg mia/conf/<YOUR_CONFIG>.yaml
+
+That's it ! The engine will now automatically fetch, filter, and process your FPV images. You may get a few errors specifying that some images were unable to be fetched due to permission limitations. That is normal and the engine will continue.
+
+Once all your locations have been downloaded, you will see that parquet files, images, and raw_images, have been populated in your `dataset_dir` for each location. You can now move on to getting BEVs.
+
+### 2. Getting BEVs
+Once you have the FPV parquet dataframes downloaded, you are now ready to fetch and generate the BEV smenatic maps. 
+
+Edit the documented bev options in your configuration file to suit your use case. The defaults are tuned to what we used to produce the MIA datasets and you can use them as is.
+
+Once configuration is done simply run the following from inside your docker container with working dir set to this repo:
+
+    python3.9 -m mia.bev.get_bev
+
+The data engine will now fetch, process, and save the semantic masks.
+
+You now have FPV-BEV pairs with associated metadata and camera parameters !
+
+**Note** to get satellite imagery for comparison you must first download it by toggling the store_sat option in the configuration
+
+### 3. (Optional) Visualize your data
+You can visualize a few samples using the tool `mia/misc_tools/vis_samples.py`. 
+
+From inside the container with working dir set to this repo, run:
+
+    python3.9 -m mia/misc_tools/vis_samples --dataset_dir /home/mia_dataset_release --locations <LOCATION_OF_INTEREST>
+
+If successful, the script will generate a PDF called `compare.pdf` in the pittsburgh directory. Upon openning you should see the metadata, FPVs, and BEVs of a few samples of the dataset. 
+
+
+## Downloading the MIA dataset
+Refer to [mia/dataset.md](mia/dataset.md) for instructions.
+
+## Training
+
+### Pre-train with MIA Dataset
+To pretrain using our paper configuration simply run:
+
+    python -m mapper.mapper data.split=<PATH TO SPLIT FILE> data.data_dir=<PATH TO MIA DATASET>
+
+### Finetune with NuScenes Dataset
+To finetune using NuScenes Dataset with our paper configuration, run:
+
+    python -m mapper.mapper -cn mapper_nuscenes training.checkpoint=<PATH TO PRETRAINED MODEL> data.data_dir=<PATH TO NUSCENES DATA> data.map_dir=<PATH TO GENERATED NUSCENES MAP>
+
+## Reproduction
+#### Dataset Setup
+**MIA**: Follow download instructions in [Downloading the MIA Dataset](#downloading-the-mia-dataset)
+
+**NuScenes**: Follow the data generation instructions in [Mono-Semantic-Maps](https://github.com/tom-roddick/mono-semantic-maps?tab=readme-ov-file#nuscenes). To match the newest available information, we use v1.3 of the NuScenes' map expansion pack. 
+
+**KITTI360-BEV**: Follow the KITTI360-BEV dataset instructions in [SkyEye](https://github.com/robot-learning-freiburg/SkyEye?tab=readme-ov-file#skyeye-datasets)
+
+#### Inference
+To generate MIA dataset prediction results(on test split), use:
+
+    python -m mapper.mapper data.split=<PATH TO SPLIT FILE> data.data_dir=<PATH TO MIA DATASET> training.checkpoint=<TRAINED WEIGHTS> training.eval=true
+*To specify location, add `data.scenes` in the argument. For example, for held-out cities `data.scenes="[pittsburgh, houston]"`*
+
+To Generate NuScenes dataset prediction results(on validation split), use:
+
+    python -m mapper.mapper -cn mapper_nuscenes training.checkpoint=<PATH TO PRETRAINED MODEL> data.data_dir=<PATH TO NUSCENES DATA> data.map_dir=<PATH TO GENERATED NUSCENES MAP> training.eval=true
+
+To Generate KITTI360-BEV dataset prediction results (on validation split), use:
+
+    python -m mapper.mapper -cn mapper_kitti training.checkpoint=<PATH TO PRETRAINED MODEL> data.seam_root_dir=<PATH TO SEAM ROOT> data.dataset_root_dir=<PATH TO KITTI DATASET> training.eval=true
+
+
+## License
+[More Information Needed]
+
+## Acknowledgement
+We thank the authors of the following repositories for their open-source code:
+- [OrienterNet](https://github.com/facebookresearch/OrienterNet)
+- [Map Machine](https://github.com/enzet/map-machine)
+- [Mono-Semantic-Maps](https://github.com/tom-roddick/mono-semantic-maps)
+- [Translating Images Into Maps](https://github.com/avishkarsaha/translating-images-into-maps)
+- [SkyEye](https://github.com/robot-learning-freiburg/SkyEye)

mapper/__init__.py

@@ -0,0 +1,30 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+import os, sys
+
+sys.path.append(os.path.dirname(os.path.realpath(__file__)))
+from pathlib import Path
+import logging
+
+import pytorch_lightning  # noqa: F401
+
+
+formatter = logging.Formatter(
+    fmt="[%(asctime)s %(name)s %(levelname)s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+handler = logging.StreamHandler()
+handler.setFormatter(formatter)
+handler.setLevel(logging.INFO)
+
+logger = logging.getLogger("mapper")
+logger.setLevel(logging.INFO)
+logger.addHandler(handler)
+logger.propagate = False
+
+pl_logger = logging.getLogger("pytorch_lightning")
+if len(pl_logger.handlers):
+    pl_logger.handlers[0].setFormatter(formatter)
+
+repo_dir = Path(__file__).parent.parent
+EXPERIMENTS_PATH = repo_dir / "experiments/"
+DATASETS_PATH = repo_dir / "datasets/"

mapper/callbacks.py

@@ -0,0 +1,105 @@
+import torch
+import pytorch_lightning as pl
+from pathlib import Path
+from typing import Any
+import torchvision
+import wandb
+
+
+class EvalSaveCallback(pl.Callback):
+
+    def __init__(self, save_dir: Path) -> None:
+        super().__init__()
+        self.save_dir = save_dir
+
+    def save(self, outputs, batch, batch_idx):
+        name = batch['name']
+
+        filename = self.save_dir / f"{batch_idx:06d}_{name[0]}.pt"
+        torch.save({
+            "fpv": batch['image'],
+            "seg_masks": batch['seg_masks'],
+            'name': name,
+            "output": outputs["output"],
+            "valid_bev": outputs["valid_bev"],
+        }, filename)
+
+    def on_test_batch_end(self, trainer: pl.Trainer,
+                          pl_module: pl.LightningModule,
+                          outputs: torch.Tensor | Any | None,
+                          batch: Any,
+                          batch_idx: int,
+                          dataloader_idx: int = 0) -> None:
+        if not outputs:
+            return
+
+        self.save(outputs, batch, batch_idx)
+
+    def on_validation_batch_end(self, trainer: pl.Trainer,
+                                pl_module: pl.LightningModule,
+                                outputs: torch.Tensor | Any | None,
+                                batch: Any,
+                                batch_idx: int,
+                                dataloader_idx: int = 0) -> None:
+        if not outputs:
+
+            return
+
+        self.save(outputs, batch, batch_idx)
+
+
+class ImageLoggerCallback(pl.Callback):
+    def __init__(self, num_classes):
+        super().__init__()
+        self.num_classes = num_classes
+
+    def log_image(self, trainer, pl_module, outputs, batch, batch_idx, mode="train"):
+        fpv_rgb = batch["image"]
+        fpv_grid = torchvision.utils.make_grid(
+            fpv_rgb, nrow=8, normalize=False)
+        images = [
+            wandb.Image(fpv_grid, caption="fpv")
+        ]
+
+        pred = outputs['output'].permute(0, 2, 3, 1)
+        pred[outputs["valid_bev"][..., :-1] == 0] = 0
+        pred = (pred > 0.5).float()
+        pred = pred.permute(0, 3, 1, 2)
+
+        for i in range(self.num_classes):
+            gt_class_i = batch['seg_masks'][..., i]
+            gt_class_i_grid = torchvision.utils.make_grid(
+                gt_class_i.unsqueeze(1), nrow=8, normalize=False, pad_value=0)
+            pred_class_i = pred[:, i]
+            pred_class_i_grid = torchvision.utils.make_grid(
+                pred_class_i.unsqueeze(1), nrow=8, normalize=False, pad_value=0)
+
+            images += [
+                wandb.Image(gt_class_i_grid, caption=f"gt_class_{i}"),
+                wandb.Image(pred_class_i_grid, caption=f"pred_class_{i}")
+            ]
+
+        trainer.logger.experiment.log(
+            {
+                "{}/images".format(mode): images
+            }
+        )
+
+    def on_validation_batch_end(self, trainer, pl_module: pl.LightningModule, outputs, batch, batch_idx):
+        if batch_idx == 0:
+            with torch.no_grad():
+                outputs = pl_module(batch)
+            self.log_image(trainer, pl_module, outputs,
+                           batch, batch_idx, mode="val")
+
+    def on_train_batch_end(self, trainer, pl_module: pl.LightningModule, outputs, batch, batch_idx):
+        if batch_idx == 0:
+            pl_module.eval()
+
+            with torch.no_grad():
+                outputs = pl_module(batch)
+
+            self.log_image(trainer, pl_module, outputs,
+                           batch, batch_idx, mode="train")
+
+            pl_module.train()

mapper/conf/data/kitti.yaml

@@ -0,0 +1,40 @@
+name: kitti
+seam_root_dir: /path/to/generated/seam
+dataset_root_dir: /path/to/kitti/dataset
+bev_percentage: 100
+pixel_per_meter: 2
+crop_size_meters: 50
+target_focal_length: 256
+resize_image: null
+pad_to_multiple: 14
+num_classes: 8
+loading:
+  train:
+    batch_size: 32
+    num_workers: 32
+  val:
+    batch_size: 32
+    num_workers: 32
+  test:
+    batch_size: 32
+    num_workers: 32
+pad_to_square: true
+rectify_pitch: true
+gravity_align: false
+class_mapping: [0, 0, 1, 2, 0, 3]
+augmentations:
+    enabled: True
+    brightness: 0.5
+    contrast: 0.5
+    saturation: 0.5
+    random_flip: 0.5
+    hue: 0.5
+    random_resized_crop: False  
+    gaussian_noise:
+        enabled: False
+        mean: 0.0
+        std: 0.1
+    brightness_contrast:
+        enabled: True
+        brightness_factor: 0.2
+        contrast_factor: 0.2

mapper/conf/data/mia.yaml

@@ -0,0 +1,44 @@
+name: mapillary
+scenes:
+- chicago
+- new_york
+- los_angeles
+- san_francisco
+split: /path/to/split/file
+data_dir: /path/to/mia/dataset
+loading:
+  train:
+    batch_size: 128
+    num_workers: 30
+  val:
+    batch_size: 128
+    num_workers: 30
+  test:
+    batch_size: 1
+    num_workers: 0
+  testsmall:
+    batch_size: 1
+    num_workers: 0
+num_classes: 6
+pixel_per_meter: 2
+crop_size_meters: 64
+resize_image: 512
+pad_to_square: true
+rectify_pitch: true
+gravity_align: true
+augmentations:
+    enabled: True
+    brightness: 0.5
+    contrast: 0.5
+    saturation: 0.5
+    random_flip: 0.5
+    hue: 0.5
+    random_resized_crop: False  
+    gaussian_noise:
+        enabled: False
+        mean: 0.0
+        std: 0.1
+    brightness_contrast:
+        enabled: True
+        brightness_factor: 0.2
+        contrast_factor: 0.2

mapper/conf/data/nuscenes.yaml

@@ -0,0 +1,38 @@
+name: nuscenes
+data_dir: /path/to/nuscenes/data
+map_dir: /path/to/generated/maps
+version: v1.0-trainval
+pixel_per_meter: 2
+crop_size_meters: 50
+resize_image: 512
+percentage: 1.0
+class_mapping: [0, 1, 2, 0, 0, 3]
+num_classes: 14
+loading:
+  train:
+    batch_size: 128
+    num_workers: 10
+  val:
+    batch_size: 128
+    num_workers: 10
+  test:
+    batch_size: 128
+    num_workers: 10
+pad_to_square: true
+rectify_pitch: true
+gravity_align: true
+augmentations:
+    enabled: True
+    brightness: 0.5
+    contrast: 0.5
+    saturation: 0.5
+    hue: 0.5
+    random_resized_crop: False  
+    gaussian_noise:
+        enabled: False
+        mean: 0.0
+        std: 0.1
+    brightness_contrast:
+        enabled: True
+        brightness_factor: 0.2
+        contrast_factor: 0.2

mapper/conf/mapper_kitti.yaml

@@ -0,0 +1,23 @@
+defaults:
+  - schema/data: kitti
+  - data: kitti
+  - model: mapper
+  - training
+  - _self_
+
+experiment:
+  name: MIA_DINOv2_Mapper_KITTI
+
+model:
+  loss:
+    xent_weight: 1.0
+    dice_weight: 1.0
+    focal_loss: false
+    focal_loss_gamma: 2.0
+    requires_frustrum: true
+    requires_flood_mask: true
+    class_weights: null
+    label_smoothing: 0.1
+
+training:
+  checkpoint: /path/to/checkpoint

mapper/conf/mapper_nuscenes.yaml

@@ -0,0 +1,26 @@
+defaults:
+  - schema/data: nuscenes
+  - data: nuscenes
+  - model: mapper
+  - training
+  - _self_
+
+experiment:
+  name: MIA_DINOv2_Mapper_NuScenes
+
+model:
+  loss:
+    xent_weight: 1.0
+    dice_weight: 1.0
+    focal_loss: false
+    focal_loss_gamma: 2.0
+    class_weights: [1.00060036,  1.85908161,  1.0249052, 0., 0., 2.57267816]
+    requires_frustrum: true
+    label_smoothing: 0.1
+
+training:
+  checkpoint: /path/to/checkpoint
+  finetune: true
+  lr: 0.0001
+  trainer:
+    max_epochs: 50

mapper/conf/model/image_encoder/dino.yaml

@@ -0,0 +1,5 @@
+name: feature_extractor_DPT
+backbone:
+  pretrained: true
+  frozen: true
+  output_dim: ${model.latent_dim} # Match Latent Dimension

mapper/conf/model/image_encoder/resnet.yaml

@@ -0,0 +1,12 @@
+name: feature_extractor_resnet
+backbone:
+  pretrained: true
+  frozen: true
+  output_dim: ${model.latent_dim} # Match Latent Dimension
+  input_dim: 3
+  encoder: resnet50
+  num_downsample: null
+  remove_stride_from_first_conv: false
+  decoder_norm: "nn.BatchNorm2d"
+  do_average_pooling: false
+  checkpointed: false

mapper/conf/model/mapper.yaml

@@ -0,0 +1,15 @@
+defaults:
+  - schema/backbone: dino
+  - image_encoder: dino
+
+segmentation_head:
+  dropout_rate: 0.2
+name: map_perception_net
+num_classes: 6
+latent_dim: 128
+z_max: 50
+x_max: 25
+pixel_per_meter: ${data.pixel_per_meter}
+num_scale_bins: 32
+loss:
+  num_classes: ${..num_classes}

mapper/conf/pretrain.yaml

@@ -0,0 +1,24 @@
+defaults:
+  - schema/data: mia
+  - data: mia
+  - model: mapper
+  - training
+  - _self_
+
+experiment:
+  name: MIA_DINOv2_Pretrain
+
+model:
+  loss:
+    xent_weight: 1.0
+    dice_weight: 1.0
+    focal_loss: false
+    focal_loss_gamma: 2.0
+    requires_frustrum: true
+    class_weights: [ 1.00351229,  4.34782609,  1.00110121, 1.03124678,
+        6.69792364,  7.55857899 ]
+    label_smoothing: 0.1
+
+training:
+  trainer:
+    max_epochs: 15

mapper/conf/pretrain_resnet.yaml

@@ -0,0 +1,26 @@
+defaults:
+  - schema/data: mia
+  - data: mia
+  - model: mapper
+  - training
+  - _self_
+  - override model/schema/backbone: resnet
+  - override model/image_encoder: resnet
+
+experiment:
+  name: MIA_DINOv2_Pretrain
+
+model:
+  loss:
+    xent_weight: 1.0
+    dice_weight: 1.0
+    focal_loss: false
+    focal_loss_gamma: 2.0
+    requires_frustrum: true
+    class_weights: [ 1.00351229,  4.34782609,  1.00110121, 1.03124678,
+        6.69792364,  7.55857899 ]
+
+training:
+  trainer:
+    max_steps: 10
+    max_epochs: 15

mapper/conf/training.yaml

@@ -0,0 +1,30 @@
+experiment:
+  name: MGL_DINOv2_v4-baseline-less-class
+  seed: 42
+training:
+  num_classes: ${model.num_classes}
+  lr: 0.001
+  lr_scheduler: 
+    name: "CosineAnnealingLR"
+    args:
+      T_max: $total_epochs
+      eta_min: 0.0000001
+  checkpoint: null
+  finetune: false
+  eval: false
+  save_dir: eval_results
+  trainer:
+    # val_check_interval: 250
+    # log_every_n_steps: 100
+    # limit_val_batches: 0
+    # max_steps: 500000
+    # num_epochs: 15
+    precision: bf16-mixed
+    accelerator: gpu
+    strategy: ddp_find_unused_parameters_true
+  checkpointing:
+    dirpath: checkpoints/
+    monitor: val/total/loss
+    save_top_k: -1
+    mode: min
+    save_last: True

mapper/data/__init__.py

@@ -0,0 +1,7 @@
+from .mapillary.data_module import MapillaryDataModule
+from .nuscenes.data_module import NuScenesData
+
+modules = {
+    "mapillary": MapillaryDataModule,
+    "nuscenes": NuScenesData
+}

mapper/data/base.py

@@ -0,0 +1,19 @@
+from abc import abstractmethod
+from typing import Optional
+
+
+class DataBase():
+    def __init__(self) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def prepare_data(self) -> None:
+        raise NotImplementedError
+    
+    @abstractmethod
+    def setup(self, stage: Optional[str] = None):
+        raise NotImplementedError
+
+    @abstractmethod
+    def dataset(self, stage: str):
+        raise NotImplementedError

mapper/data/image.py

@@ -0,0 +1,140 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+from typing import Callable, Optional, Union, Sequence
+
+import numpy as np
+import torch
+import torchvision.transforms.functional as tvf
+import collections
+from scipy.spatial.transform import Rotation
+
+from ..utils.geometry import from_homogeneous, to_homogeneous
+from ..utils.wrappers import Camera
+
+
+def rectify_image(
+    image: torch.Tensor,
+    cam: Camera,
+    roll: float,
+    pitch: Optional[float] = None,
+    valid: Optional[torch.Tensor] = None,
+):
+    *_, h, w = image.shape
+    grid = torch.meshgrid(
+        [torch.arange(w, device=image.device), torch.arange(h, device=image.device)],
+        indexing="xy",
+    )
+    grid = torch.stack(grid, -1).to(image.dtype)
+
+    if pitch is not None:
+        args = ("ZX", (roll, pitch))
+    else:
+        args = ("Z", roll)
+    R = Rotation.from_euler(*args, degrees=True).as_matrix()
+    R = torch.from_numpy(R).to(image)
+
+    grid_rect = to_homogeneous(cam.normalize(grid)) @ R.T
+    grid_rect = cam.denormalize(from_homogeneous(grid_rect))
+    grid_norm = (grid_rect + 0.5) / grid.new_tensor([w, h]) * 2 - 1
+    rectified = torch.nn.functional.grid_sample(
+        image[None],
+        grid_norm[None],
+        align_corners=False,
+        mode="bilinear",
+    ).squeeze(0)
+    if valid is None:
+        valid = torch.all((grid_norm >= -1) & (grid_norm <= 1), -1)
+    else:
+        valid = (
+            torch.nn.functional.grid_sample(
+                valid[None, None].float(),
+                grid_norm[None],
+                align_corners=False,
+                mode="nearest",
+            )[0, 0]
+            > 0
+        )
+    return rectified, valid
+
+
+def resize_image(
+    image: torch.Tensor,
+    size: Union[int, Sequence, np.ndarray],
+    fn: Optional[Callable] = None,
+    camera: Optional[Camera] = None,
+    valid: np.ndarray = None,
+):
+    """Resize an image to a fixed size, or according to max or min edge."""
+    *_, h, w = image.shape
+    if fn is not None:
+        assert isinstance(size, int)
+        scale = size / fn(h, w)
+        h_new, w_new = int(round(h * scale)), int(round(w * scale))
+        scale = (scale, scale)
+    else:
+        if isinstance(size, (collections.abc.Sequence, np.ndarray)):
+            w_new, h_new = size
+        elif isinstance(size, int):
+            w_new = h_new = size
+        else:
+            raise ValueError(f"Incorrect new size: {size}")
+        scale = (w_new / w, h_new / h)
+    if (w, h) != (w_new, h_new):
+        mode = tvf.InterpolationMode.BILINEAR
+        image = tvf.resize(image, (int(h_new), int(w_new)), interpolation=mode, antialias=True)
+        image.clip_(0, 1)
+        if camera is not None:
+            camera = camera.scale(scale)
+        if valid is not None:
+            valid = tvf.resize(
+                valid.unsqueeze(0),
+                (int(h_new), int(w_new)),
+                interpolation=tvf.InterpolationMode.NEAREST,
+            ).squeeze(0)
+    ret = [image, scale]
+    if camera is not None:
+        ret.append(camera)
+    if valid is not None:
+        ret.append(valid)
+    return ret
+
+
+def pad_image(
+    image: torch.Tensor,
+    size: Union[int, Sequence, np.ndarray],
+    camera: Optional[Camera] = None,
+    valid: torch.Tensor = None,
+    crop_and_center: bool = False,
+):
+    if isinstance(size, int):
+        w_new = h_new = size
+    elif isinstance(size, (collections.abc.Sequence, np.ndarray)):
+        w_new, h_new = size
+    else:
+        raise ValueError(f"Incorrect new size: {size}")
+    *c, h, w = image.shape
+    if crop_and_center:
+        diff = np.array([w - w_new, h - h_new])
+        left, top = left_top = np.round(diff / 2).astype(int)
+        right, bottom = diff - left_top
+    else:
+        assert h <= h_new
+        assert w <= w_new
+        top = bottom = left = right = 0
+    slice_out = np.s_[..., : min(h, h_new), : min(w, w_new)]
+    slice_in = np.s_[
+        ..., max(top, 0) : h - max(bottom, 0), max(left, 0) : w - max(right, 0)
+    ]
+    if (w, h) == (w_new, h_new):
+        out = image
+    else:
+        out = torch.zeros((*c, h_new, w_new), dtype=image.dtype)
+        out[slice_out] = image[slice_in]
+        if camera is not None:
+            camera = camera.crop((max(left, 0), max(top, 0)), (w_new, h_new))
+    out_valid = torch.zeros((h_new, w_new), dtype=torch.bool)
+    out_valid[slice_out] = True if valid is None else valid[slice_in]
+    if camera is not None:
+        return out, out_valid, camera
+    else:
+        return out, out_valid

mapper/data/kitti/data_module.py

@@ -0,0 +1,32 @@
+from ..base import DataBase
+from .dataset import BEVKitti360Dataset
+from ..schema import KITTIDataConfiguration
+
+class BEVKitti360Data(DataBase):
+    def __init__(self, cfg: KITTIDataConfiguration) -> None:
+        self.cfg = cfg
+        self._dataset = {}
+
+    def prepare_data(self) -> None:
+        return
+    
+    def setup(self, stage: str) -> None:
+        split = {
+            'fit': 'train',
+            'val': 'val',
+            'validate': 'val',
+            'test': 'val',
+            "train": "train"
+        }[stage]
+
+        self._dataset[stage] = BEVKitti360Dataset(
+            cfg=self.cfg,
+            split_name=split
+        )
+
+    def dataset(self, stage: str):
+        if self._dataset.get(stage) is None:
+            self.setup(stage)
+        
+        return self._dataset[stage]
+    

mapper/data/kitti/dataset.py

@@ -0,0 +1,317 @@
+import os
+import numpy as np
+import torch.utils.data as data
+import umsgpack
+from PIL import Image
+import json
+import torchvision.transforms as tvf
+
+from .transform import BEVTransform
+from ..schema import KITTIDataConfiguration
+
+class BEVKitti360Dataset(data.Dataset):
+    _IMG_DIR = "img"
+    _BEV_MSK_DIR = "bev_msk"
+    _BEV_PLABEL_DIR = "bev_plabel_dynamic"
+    _FV_MSK_DIR = "front_msk_seam"
+    _BEV_DIR = "bev_ortho"
+    _LST_DIR = "split"
+    _PERCENTAGES_DIR = "percentages"
+    _BEV_METADATA_FILE = "metadata_ortho.bin"
+    _FV_METADATA_FILE = "metadata_front.bin"
+
+    def __init__(self, cfg: KITTIDataConfiguration, split_name="train"):
+        super(BEVKitti360Dataset, self).__init__()
+        self.cfg = cfg
+        self.seam_root_dir = cfg.seam_root_dir  # Directory of seamless data
+        self.kitti_root_dir = cfg.dataset_root_dir  #  Directory of the KITTI360 data
+        self.split_name = split_name
+        
+        self.rgb_cameras = ['front']
+        if cfg.bev_percentage < 1:
+            self.bev_percentage = cfg.bev_percentage
+        else:
+            self.bev_percentage = int(cfg.bev_percentage)
+
+        # Folders
+        self._img_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._IMG_DIR)
+        self._bev_msk_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._BEV_MSK_DIR, BEVKitti360Dataset._BEV_DIR)
+        self._bev_plabel_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._BEV_PLABEL_DIR, BEVKitti360Dataset._BEV_DIR)
+        self._fv_msk_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._FV_MSK_DIR, "front")
+        self._lst_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._LST_DIR)
+        self._percentages_dir = os.path.join(self.seam_root_dir, BEVKitti360Dataset._LST_DIR, BEVKitti360Dataset._PERCENTAGES_DIR)
+
+        # Load meta-data and split
+        self._bev_meta, self._bev_images, self._bev_images_all, self._fv_meta, self._fv_images, self._fv_images_all,\
+        self._img_map, self.bev_percent_split = self._load_split()
+
+        self.tfs = self.get_augmentations() if split_name == "train" else tvf.Compose([])
+        self.transform = BEVTransform(cfg, self.tfs)
+
+    def get_augmentations(self):
+
+        print(f"Augmentation!", "\n" * 10)
+        augmentations = [
+            tvf.ColorJitter(
+                brightness=self.cfg.augmentations.brightness,
+                contrast=self.cfg.augmentations.contrast,
+                saturation=self.cfg.augmentations.saturation,
+                hue=self.cfg.augmentations.hue,
+            )
+        ]
+
+        if self.cfg.augmentations.random_resized_crop:
+            augmentations.append(
+                tvf.RandomResizedCrop(scale=(0.8, 1.0))
+            )  # RandomResizedCrop
+
+        if self.cfg.augmentations.gaussian_noise.enabled:
+            augmentations.append(
+                tvf.GaussianNoise(
+                    mean=self.cfg.augmentations.gaussian_noise.mean,
+                    std=self.cfg.augmentations.gaussian_noise.std,
+                )
+            )  # Gaussian noise
+
+        if self.cfg.augmentations.brightness_contrast.enabled:
+            augmentations.append(
+                tvf.ColorJitter(
+                    brightness=self.cfg.augmentations.brightness_contrast.brightness_factor,
+                    contrast=self.cfg.augmentations.brightness_contrast.contrast_factor,
+                    saturation=0,  # Keep saturation at 0 for brightness and contrast adjustment
+                    hue=0,
+                )
+            )  # Brightness and contrast adjustment
+
+        return tvf.Compose(augmentations)
+    
+    # Load the train or the validation split
+    def _load_split(self):
+        with open(os.path.join(self.seam_root_dir, BEVKitti360Dataset._BEV_METADATA_FILE), "rb") as fid:
+            bev_metadata = umsgpack.unpack(fid, encoding="utf-8")
+
+        with open(os.path.join(self.seam_root_dir, BEVKitti360Dataset._FV_METADATA_FILE), 'rb') as fid:
+            fv_metadata = umsgpack.unpack(fid, encoding="utf-8")
+
+        # Read the files for this split
+        with open(os.path.join(self._lst_dir, self.split_name + ".txt"), "r") as fid:
+            lst = fid.readlines()
+            lst = [line.strip() for line in lst]
+
+        if self.split_name == "train":
+            # Get all the frames in the train dataset. This will be used for generating samples for temporal consistency.
+            with open(os.path.join(self._lst_dir, "{}_all.txt".format(self.split_name)), 'r') as fid:
+                lst_all = fid.readlines()
+                lst_all = [line.strip() for line in lst_all]
+
+            # Get all the samples for which the BEV plabels have to be loaded.
+            percentage_file = os.path.join(self._percentages_dir, "{}_{}.txt".format(self.split_name, self.bev_percentage))
+            print("Loading {}% file".format(self.bev_percentage))
+            with open(percentage_file, 'r') as fid:
+                lst_percent = fid.readlines()
+                lst_percent = [line.strip() for line in lst_percent]
+        else:
+            lst_all = lst
+            lst_percent = lst
+
+        # Remove elements from lst if they are not in _FRONT_MSK_DIR
+        fv_msk_frames = os.listdir(self._fv_msk_dir)
+        fv_msk_frames = [frame.split(".")[0] for frame in fv_msk_frames]
+        fv_msk_frames_exist_map = {entry: True for entry in fv_msk_frames}  # This is to speed-up the dataloader
+        lst = [entry for entry in lst if entry in fv_msk_frames_exist_map]
+        lst_all = [entry for entry in lst_all if entry in fv_msk_frames_exist_map]
+
+        # Filter based on the samples plabels
+        if self.bev_percentage < 100:
+            lst_filt = [entry for entry in lst if entry in lst_percent]
+            lst = lst_filt
+
+        # Remove any potential duplicates
+        lst = set(lst)
+        lst_percent = set(lst_percent)
+
+        img_map = {}
+        for camera in self.rgb_cameras:
+            with open(os.path.join(self._img_dir, "{}.json".format(camera))) as fp:
+                map_list = json.load(fp)
+                map_dict = {k: v for d in map_list for k, v in d.items()}
+                img_map[camera] = map_dict
+
+        bev_meta = bev_metadata["meta"]
+        bev_images = [img_desc for img_desc in bev_metadata["images"] if img_desc["id"] in lst]
+        fv_meta = fv_metadata["meta"]
+        fv_images = [img_desc for img_desc in fv_metadata['images'] if img_desc['id'] in lst]
+
+        # Check for inconsistency due to inconsistencies in the input files or dataset
+        bev_images_ids = [bev_img["id"] for bev_img in bev_images]
+        fv_images_ids = [fv_img["id"] for fv_img in fv_images]
+        assert set(bev_images_ids) == set(fv_images_ids) and len(bev_images_ids) == len(fv_images_ids), 'Inconsistency between fv_images and bev_images detected'
+
+        if lst_all is not None:
+            bev_images_all = [img_desc for img_desc in bev_metadata['images'] if img_desc['id'] in lst_all]
+            fv_images_all = [img_desc for img_desc in fv_metadata['images'] if img_desc['id'] in lst_all]
+        else:
+            bev_images_all, fv_images_all = None, None
+
+        return bev_meta, bev_images, bev_images_all, fv_meta, fv_images, fv_images_all, img_map, lst_percent
+
+    def _find_index(self, list, key, value):
+        for i, dic in enumerate(list):
+            if dic[key] == value:
+                return i
+        return None
+
+    def _load_item(self, item_idx):
+        # Find the index of the element in the list containing all elements
+        all_idx = self._find_index(self._fv_images_all, "id", self._fv_images[item_idx]['id'])
+        if all_idx is None:
+            raise IOError("Required index not found!")
+        
+        bev_img_desc = self._bev_images[item_idx]
+        fv_img_desc = self._fv_images[item_idx]
+
+        scene, frame_id = self._bev_images[item_idx]["id"].split(";")
+
+        # Get the RGB file names
+        img_file = os.path.join(
+            self.kitti_root_dir,
+            self._img_map["front"]["{}.png"
+                                   .format(bev_img_desc['id'])]
+        )
+
+        if not os.path.exists(img_file):
+            raise IOError(
+                "RGB image not found! Scene: {}, Frame: {}".format(scene, frame_id)
+            )
+
+        # Load the images
+        img = Image.open(img_file).convert(mode="RGB")
+
+        # Load the BEV mask
+        bev_msk_file = os.path.join(
+            self._bev_msk_dir,
+            "{}.png".format(bev_img_desc['id'])
+        )
+        bev_msk = Image.open(bev_msk_file)
+        bev_plabel = None
+
+        # Load the front mask
+        fv_msk_file = os.path.join(
+            self._fv_msk_dir, 
+            "{}.png".format(fv_img_desc['id'])
+        )
+        fv_msk = Image.open(fv_msk_file)
+
+
+        bev_weights_msk_combined = None
+
+        # Get the other information
+        bev_cat = bev_img_desc["cat"]
+        bev_iscrowd = bev_img_desc["iscrowd"]
+        fv_cat = fv_img_desc['cat']
+        fv_iscrowd = fv_img_desc['iscrowd']
+        fv_intrinsics = fv_img_desc["cam_intrinsic"]
+        ego_pose = fv_img_desc['ego_pose']  # This loads the cam0 pose
+
+        # Get the ids of all the frames
+        frame_ids = bev_img_desc["id"]
+
+        return img, bev_msk, bev_plabel, fv_msk, bev_weights_msk_combined, bev_cat, \
+            bev_iscrowd, fv_cat, fv_iscrowd, fv_intrinsics, ego_pose, frame_ids
+
+    @property
+    def fv_categories(self):
+        """Category names"""
+        return self._fv_meta["categories"]
+
+    @property
+    def fv_num_categories(self):
+        """Number of categories"""
+        return len(self.fv_categories)
+
+    @property
+    def fv_num_stuff(self):
+        """Number of "stuff" categories"""
+        return self._fv_meta["num_stuff"]
+
+    @property
+    def fv_num_thing(self):
+        """Number of "thing" categories"""
+        return self.fv_num_categories - self.fv_num_stuff
+
+    @property
+    def bev_categories(self):
+        """Category names"""
+        return self._bev_meta["categories"]
+
+    @property
+    def bev_num_categories(self):
+        """Number of categories"""
+        return len(self.bev_categories)
+
+    @property
+    def bev_num_stuff(self):
+        """Number of "stuff" categories"""
+        return self._bev_meta["num_stuff"]
+
+    @property
+    def bev_num_thing(self):
+        """Number of "thing" categories"""
+        return self.bev_num_categories - self.bev_num_stuff
+
+    @property
+    def original_ids(self):
+        """Original class id of each category"""
+        return self._fv_meta["original_ids"]
+
+    @property
+    def palette(self):
+        """Default palette to be used when color-coding semantic labels"""
+        return np.array(self._fv_meta["palette"], dtype=np.uint8)
+
+    @property
+    def img_sizes(self):
+        """Size of each image of the dataset"""
+        return [img_desc["size"] for img_desc in self._fv_images]
+
+    @property
+    def img_categories(self):
+        """Categories present in each image of the dataset"""
+        return [img_desc["cat"] for img_desc in self._fv_images]
+
+    @property
+    def dataset_name(self):
+        return "Kitti360"
+
+    def __len__(self):
+        if self.cfg.percentage < 1:
+            return int(len(self._fv_images) * self.cfg.percentage)
+        
+        return len(self._fv_images)
+
+    def __getitem__(self, item):
+        img, bev_msk, bev_plabel, fv_msk, bev_weights_msk, bev_cat, bev_iscrowd, fv_cat, fv_iscrowd, fv_intrinsics, ego_pose, idx = self._load_item(item)
+        
+        rec = self.transform(img=img, bev_msk=bev_msk, bev_plabel=bev_plabel, fv_msk=fv_msk, bev_weights_msk=bev_weights_msk, bev_cat=bev_cat,
+                             bev_iscrowd=bev_iscrowd, fv_cat=fv_cat, fv_iscrowd=fv_iscrowd, fv_intrinsics=fv_intrinsics,
+                             ego_pose=ego_pose)
+        size = (img.size[1], img.size[0])
+
+        # Close the file
+        img.close()
+        bev_msk.close()
+        fv_msk.close()
+
+        rec["index"] = idx
+        rec["size"] = size
+        rec['name'] = idx 
+        
+        return rec
+
+    def get_image_desc(self, idx):
+        """Look up an image descriptor given the id"""
+        matching = [img_desc for img_desc in self._images if img_desc["id"] == idx]
+        if len(matching) == 1:
+            return matching[0]
+        else:
+            raise ValueError("No image found with id %s" % idx)

mapper/data/kitti/transform.py

@@ -0,0 +1,149 @@
+import numpy as np
+import torch
+from torchvision.transforms import functional as tfn
+import torchvision.transforms.functional as tvf
+
+from ..utils import decompose_rotmat
+from ..image import pad_image, rectify_image, resize_image
+from ...utils.wrappers import Camera
+from ..schema import KITTIDataConfiguration
+
+
+class BEVTransform:
+    def __init__(self,
+                 cfg: KITTIDataConfiguration, augmentations):
+        self.cfg = cfg
+        self.augmentations = augmentations
+
+    @staticmethod
+    def _compact_labels(msk, cat, iscrowd):
+        ids = np.unique(msk)
+        if 0 not in ids:
+            ids = np.concatenate((np.array([0], dtype=np.int32), ids), axis=0)
+
+        ids_to_compact = np.zeros((ids.max() + 1,), dtype=np.int32)
+        ids_to_compact[ids] = np.arange(0, ids.size, dtype=np.int32)
+
+        msk = ids_to_compact[msk]
+        cat = cat[ids]
+        iscrowd = iscrowd[ids]
+
+        return msk, cat, iscrowd
+
+    def __call__(self, img, bev_msk=None, bev_plabel=None, fv_msk=None, bev_weights_msk=None,
+                 bev_cat=None, bev_iscrowd=None, fv_cat=None, fv_iscrowd=None,
+                 fv_intrinsics=None, ego_pose=None):
+        # Wrap in np.array
+        if bev_cat is not None:
+            bev_cat = np.array(bev_cat, dtype=np.int32)
+        if bev_iscrowd is not None:
+            bev_iscrowd = np.array(bev_iscrowd, dtype=np.uint8)
+
+        if ego_pose is not None:
+            ego_pose = np.array(ego_pose, dtype=np.float32)
+
+        roll, pitch, yaw = decompose_rotmat(ego_pose[:3, :3])
+
+        # Image transformations
+        img = tfn.to_tensor(img)
+        # img = [self._normalize_image(rgb) for rgb in img]
+        fx = fv_intrinsics[0][0]
+        fy = fv_intrinsics[1][1]
+        cx = fv_intrinsics[0][2]
+        cy = fv_intrinsics[1][2]
+        width = img.shape[2]
+        height = img.shape[1]
+
+        cam = Camera(torch.tensor(
+                        [width, height, fx, fy, cx - 0.5, cy - 0.5])).float()
+
+        if not self.cfg.gravity_align:
+            # Turn off gravity alignment
+            roll = 0.0
+            pitch = 0.0
+            img, valid = rectify_image(img, cam, roll, pitch)
+        else:
+            img, valid = rectify_image(
+                img, cam, roll, pitch if self.cfg.rectify_pitch else None
+            )
+            roll = 0.0
+            if self.cfg.rectify_pitch:
+                pitch = 0.0
+
+        if self.cfg.target_focal_length is not None:
+            # Resize to a canonical focal length
+            factor = self.cfg.target_focal_length / cam.f.numpy()
+            size = (np.array(img.shape[-2:][::-1]) * factor).astype(int)
+            img, _, cam, valid = resize_image(img, size, camera=cam, valid=valid)
+            size_out = self.cfg.resize_image
+            if size_out is None:
+                # Round the edges up such that they are multiple of a factor
+                stride = self.cfg.pad_to_multiple
+                size_out = (np.ceil((size / stride)) * stride).astype(int)
+            # Crop or pad such that both edges are of the given size
+            img, valid, cam = pad_image(
+                img, size_out, cam, valid, crop_and_center=False
+            )
+        elif self.cfg.resize_image is not None:
+            img, _, cam, valid = resize_image(
+                img, self.cfg.resize_image, fn=max, camera=cam, valid=valid
+            )
+            if self.cfg.pad_to_square:
+                # Pad such that both edges are of the given size
+                img, valid, cam = pad_image(img, self.cfg.resize_image, cam, valid)
+
+        # Label transformations,
+        if bev_msk is not None:
+            bev_msk = np.expand_dims(
+                np.array(bev_msk, dtype=np.int32, copy=False),
+                axis=0
+            )
+            bev_msk, bev_cat, bev_iscrowd = self._compact_labels(
+                bev_msk, bev_cat, bev_iscrowd
+            )
+
+            bev_msk = torch.from_numpy(bev_msk)
+            bev_cat = torch.from_numpy(bev_cat)
+
+            rotated_mask = torch.rot90(bev_msk, dims=(1, 2))
+            cropped_mask = rotated_mask[:, :672, (rotated_mask.size(2) - 672) // 2:-(rotated_mask.size(2) - 672) // 2]
+
+            bev_msk = cropped_mask.squeeze(0)
+            seg_masks = bev_cat[bev_msk]
+
+            seg_masks_onehot = seg_masks.clone()
+            seg_masks_onehot[seg_masks_onehot == 255] = 0
+            seg_masks_onehot = torch.nn.functional.one_hot(
+                seg_masks_onehot.to(torch.int64),
+                num_classes=self.cfg.num_classes
+            )
+            seg_masks_onehot[seg_masks == 255] = 0
+
+            seg_masks_onehot = seg_masks_onehot.permute(2, 0, 1)
+
+            seg_masks_down = tvf.resize(seg_masks_onehot, (100, 100))
+
+            seg_masks_down = seg_masks_down.permute(1, 2, 0)
+
+            if self.cfg.class_mapping is not None:
+                seg_masks_down = seg_masks_down[:, :, self.cfg.class_mapping]
+
+        img = self.augmentations(img)
+        flood_masks = torch.all(seg_masks_down == 0, dim=2).float()
+
+
+        ret = {
+            "image": img,
+            "valid": valid,
+            "camera": cam,
+            "seg_masks": (seg_masks_down).float().contiguous(),
+            "flood_masks": flood_masks,
+            "roll_pitch_yaw": torch.tensor((roll, pitch, yaw)).float(),
+            "confidence_map": flood_masks,
+        }
+        
+        for key, value in ret.items():
+            if isinstance(value, np.ndarray):
+                ret[key] = torch.from_numpy(value)
+    
+        return ret

mapper/data/mapillary/data_module.py

@@ -0,0 +1,317 @@
+import json
+from collections import defaultdict
+import os
+import shutil
+import tarfile
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import pytorch_lightning as pl
+import torch
+import torch.utils.data as torchdata
+from omegaconf import DictConfig
+
+from ... import logger
+from .dataset import MapLocDataset
+from ..sequential import chunk_sequence
+from ..torch import collate, worker_init_fn
+from ..schema import MIADataConfiguration
+
+def pack_dump_dict(dump):
+    for per_seq in dump.values():
+        if "points" in per_seq:
+            for chunk in list(per_seq["points"]):
+                points = per_seq["points"].pop(chunk)
+                if points is not None:
+                    per_seq["points"][chunk] = np.array(
+                        per_seq["points"][chunk], np.float64
+                    )
+        for view in per_seq["views"].values():
+            for k in ["R_c2w", "roll_pitch_yaw"]:
+                view[k] = np.array(view[k], np.float32)
+            for k in ["chunk_id"]:
+                if k in view:
+                    view.pop(k)
+        if "observations" in view:
+            view["observations"] = np.array(view["observations"])
+        for camera in per_seq["cameras"].values():
+            for k in ["params"]:
+                camera[k] = np.array(camera[k], np.float32)
+    return dump
+
+
+class MapillaryDataModule(pl.LightningDataModule):
+    dump_filename = "dump.json"
+    images_archive = "images.tar.gz"
+    images_dirname = "images/"
+    semantic_masks_dirname = "semantic_masks/"
+    flood_dirname = "flood_fill/"
+
+    def __init__(self, cfg: MIADataConfiguration):
+        super().__init__()
+        self.cfg = cfg
+        self.root = self.cfg.data_dir
+        self.local_dir = None
+
+    def prepare_data(self):
+        for scene in self.cfg.scenes:
+            dump_dir = self.root / scene
+            assert (dump_dir / self.dump_filename).exists(), dump_dir
+            # assert (dump_dir / self.cfg.tiles_filename).exists(), dump_dir
+            if self.local_dir is None:
+                assert (dump_dir / self.images_dirname).exists(), dump_dir
+                continue
+            assert (dump_dir / self.semantic_masks_dirname).exists(), dump_dir
+            assert (dump_dir / self.flood_dirname).exists(), dump_dir
+            # Cache the folder of images locally to speed up reading
+            local_dir = self.local_dir / scene
+            if local_dir.exists():
+                shutil.rmtree(local_dir)
+            local_dir.mkdir(exist_ok=True, parents=True)
+            images_archive = dump_dir / self.images_archive
+            logger.info("Extracting the image archive %s.", images_archive)
+            with tarfile.open(images_archive) as fp:
+                fp.extractall(local_dir)
+
+    def setup(self, stage: Optional[str] = None):
+        self.dumps = {}
+        # self.tile_managers = {}
+        self.image_dirs = {}
+        self.seg_masks_dir = {}
+        self.flood_masks_dir = {}
+        names = []
+
+        for scene in self.cfg.scenes:
+            logger.info("Loading scene %s.", scene)
+            dump_dir = self.root / scene
+
+            logger.info("Loading dump json file %s.", self.dump_filename)
+            with (dump_dir / self.dump_filename).open("r") as fp:
+                self.dumps[scene] = pack_dump_dict(json.load(fp))
+            for seq, per_seq in self.dumps[scene].items():
+                for cam_id, cam_dict in per_seq["cameras"].items():
+                    if cam_dict["model"] != "PINHOLE":
+                        raise ValueError(
+                            f"Unsupported camera model: {cam_dict['model']} for {scene},{seq},{cam_id}"
+                        )
+
+            self.image_dirs[scene] = (
+                (self.local_dir or self.root) / scene / self.images_dirname
+            )
+            assert self.image_dirs[scene].exists(), self.image_dirs[scene]
+
+            self.seg_masks_dir[scene] = (
+                (self.local_dir or self.root) / scene / self.semantic_masks_dirname
+            )   
+            assert self.seg_masks_dir[scene].exists(), self.seg_masks_dir[scene]
+
+            self.flood_masks_dir[scene] = (
+                (self.local_dir or self.root) / scene / self.flood_dirname
+            )   
+            assert self.flood_masks_dir[scene].exists(), self.flood_masks_dir[scene]
+
+            images = set(x.split('.')[0] for x in os.listdir(self.image_dirs[scene]))
+            flood_masks = set(x.split('.')[0] for x in os.listdir(self.flood_masks_dir[scene]))
+            semantic_masks = set(x.split('.')[0] for x in os.listdir(self.seg_masks_dir[scene]))
+
+            for seq, data in self.dumps[scene].items():
+                for name in data["views"]:
+                    if name in images and name.split("_")[0] in flood_masks and name.split("_")[0] in semantic_masks:
+                        names.append((scene, seq, name))
+ 
+        self.parse_splits(self.cfg.split, names)
+        if self.cfg.filter_for is not None:
+            self.filter_elements()
+        self.pack_data()
+
+    def pack_data(self):
+        # We pack the data into compact tensors that can be shared across processes without copy
+        exclude = {
+            "compass_angle",
+            "compass_accuracy",
+            "gps_accuracy",
+            "chunk_key",
+            "panorama_offset",
+        }
+        cameras = {
+            scene: {seq: per_seq["cameras"] for seq, per_seq in per_scene.items()}
+            for scene, per_scene in self.dumps.items()
+        }
+        points = {
+            scene: {
+                seq: {
+                    i: torch.from_numpy(p) for i, p in per_seq.get("points", {}).items()
+                }
+                for seq, per_seq in per_scene.items()
+            }
+            for scene, per_scene in self.dumps.items()
+        }
+        self.data = {}
+
+        # TODO: remove
+        if self.cfg.split == "splits_MGL_13loc.json":
+        # Use Last 20% as Val
+            num_samples_to_move = int(len(self.splits['train']) * 0.2)
+            samples_to_move = self.splits['train'][-num_samples_to_move:]
+            self.splits['val'].extend(samples_to_move)
+            self.splits['train'] = self.splits['train'][:-num_samples_to_move]
+            print(f"Dataset Len: {len(self.splits['train']), len(self.splits['val'])}\n\n\n\n")
+        elif self.cfg.split == "splits_MGL_soma_70k_mappred_random.json":
+            for stage, names in self.splits.items():
+                print("Length of splits {}: ".format(stage), len(self.splits[stage]))
+        for stage, names in self.splits.items():
+            view = self.dumps[names[0][0]][names[0][1]]["views"][names[0][2]]
+            data = {k: [] for k in view.keys() - exclude}
+            for scene, seq, name in names:
+                for k in data:
+                    data[k].append(self.dumps[scene][seq]["views"][name].get(k, None))
+            for k in data:
+                v = np.array(data[k])
+                if np.issubdtype(v.dtype, np.integer) or np.issubdtype(
+                    v.dtype, np.floating
+                ):
+                    v = torch.from_numpy(v)
+                data[k] = v
+            data["cameras"] = cameras
+            data["points"] = points
+            self.data[stage] = data
+            self.splits[stage] = np.array(names)
+
+    def filter_elements(self):
+        for stage, names in self.splits.items():
+            names_select = []
+            for scene, seq, name in names:
+                view = self.dumps[scene][seq]["views"][name]
+                if self.cfg.filter_for == "ground_plane":
+                    if not (1.0 <= view["height"] <= 3.0):
+                        continue
+                    planes = self.dumps[scene][seq].get("plane")
+                    if planes is not None:
+                        inliers = planes[str(view["chunk_id"])][-1]
+                        if inliers < 10:
+                            continue
+                    if self.cfg.filter_by_ground_angle is not None:
+                        plane = np.array(view["plane_params"])
+                        normal = plane[:3] / np.linalg.norm(plane[:3])
+                        angle = np.rad2deg(np.arccos(np.abs(normal[-1])))
+                        if angle > self.cfg.filter_by_ground_angle:
+                            continue
+                elif self.cfg.filter_for == "pointcloud":
+                    if len(view["observations"]) < self.cfg.min_num_points:
+                        continue
+                elif self.cfg.filter_for is not None:
+                    raise ValueError(f"Unknown filtering: {self.cfg.filter_for}")
+                names_select.append((scene, seq, name))
+            logger.info(
+                "%s: Keep %d/%d images after filtering for %s.",
+                stage,
+                len(names_select),
+                len(names),
+                self.cfg.filter_for,
+            )
+            self.splits[stage] = names_select
+
+    def parse_splits(self, split_arg, names):
+        if split_arg is None:
+            self.splits = {
+                "train": names,
+                "val": names,
+            }
+        elif isinstance(split_arg, int):
+            names = np.random.RandomState(self.cfg.seed).permutation(names).tolist()
+            self.splits = {
+                "train": names[split_arg:],
+                "val": names[:split_arg],
+            }
+        elif isinstance(split_arg, float):
+            names = np.random.RandomState(self.cfg.seed).permutation(names).tolist()
+            self.splits = {
+                "train": names[int(split_arg * len(names)) :],
+                "val": names[: int(split_arg * len(names))],
+            }
+        elif isinstance(split_arg, DictConfig):
+            scenes_val = set(split_arg.val)
+            scenes_train = set(split_arg.train)
+            assert len(scenes_val - set(self.cfg.scenes)) == 0
+            assert len(scenes_train - set(self.cfg.scenes)) == 0
+            self.splits = {
+                "train": [n for n in names if n[0] in scenes_train],
+                "val": [n for n in names if n[0] in scenes_val],
+            }
+        elif isinstance(split_arg, str):
+            
+            if "/" in split_arg:
+                split_path = self.root / split_arg
+            else:
+                split_path = Path(split_arg)
+            
+            with split_path.open("r") as fp:
+                splits = json.load(fp)
+            splits = {
+                k: {loc: set(ids) for loc, ids in split.items()}
+                for k, split in splits.items()
+            }
+            self.splits = {}
+            
+            for k, split in splits.items():
+                self.splits[k] = [
+                    n
+                    for n in names
+                    if n[0] in split and int(n[-1].rsplit("_", 1)[0]) in split[n[0]]
+                ]
+        else:
+            raise ValueError(split_arg)
+
+    def dataset(self, stage: str):
+        return MapLocDataset(
+            stage,
+            self.cfg,
+            self.splits[stage],
+            self.data[stage],
+            self.image_dirs,
+            self.seg_masks_dir,
+            self.flood_masks_dir,
+
+            image_ext=".jpg",
+        )
+
+    def sequence_dataset(self, stage: str, **kwargs):
+        keys = self.splits[stage]
+        seq2indices = defaultdict(list)
+        for index, (_, seq, _) in enumerate(keys):
+            seq2indices[seq].append(index)
+        # chunk the sequences to the required length
+        chunk2indices = {}
+        for seq, indices in seq2indices.items():
+            chunks = chunk_sequence(self.data[stage], indices, **kwargs)
+            for i, sub_indices in enumerate(chunks):
+                chunk2indices[seq, i] = sub_indices
+        # store the index of each chunk in its sequence
+        chunk_indices = torch.full((len(keys),), -1)
+        for (_, chunk_index), idx in chunk2indices.items():
+            chunk_indices[idx] = chunk_index
+        self.data[stage]["chunk_index"] = chunk_indices
+        dataset = self.dataset(stage)
+        return dataset, chunk2indices
+
+    def sequence_dataloader(self, stage: str, shuffle: bool = False, **kwargs):
+        dataset, chunk2idx = self.sequence_dataset(stage, **kwargs)
+        chunk_keys = sorted(chunk2idx)
+        if shuffle:
+            perm = torch.randperm(len(chunk_keys))
+            chunk_keys = [chunk_keys[i] for i in perm]
+        key_indices = [i for key in chunk_keys for i in chunk2idx[key]]
+        num_workers = self.cfg.loading[stage]["num_workers"]
+        loader = torchdata.DataLoader(
+            dataset,
+            batch_size=None,
+            sampler=key_indices,
+            num_workers=num_workers,
+            shuffle=False,
+            pin_memory=True,
+            persistent_workers=num_workers > 0,
+            worker_init_fn=worker_init_fn,
+            collate_fn=collate,
+        )
+        return loader, chunk_keys, chunk2idx

mapper/data/mapillary/dataset.py

@@ -0,0 +1,255 @@
+from copy import deepcopy
+from pathlib import Path
+from typing import Any, Dict, List
+
+import numpy as np
+import torch
+import torch.utils.data as torchdata
+import torchvision.transforms as tvf
+from PIL import Image
+from pathlib import Path
+
+from ...models.utils import deg2rad, rotmat2d
+from ...utils.io import read_image
+from ...utils.wrappers import Camera
+from ..image import pad_image, rectify_image, resize_image
+from ..utils import decompose_rotmat
+from ..schema import MIADataConfiguration
+
+
+class MapLocDataset(torchdata.Dataset):
+    def __init__(
+        self,
+        stage: str,
+        cfg: MIADataConfiguration,
+        names: List[str],
+        data: Dict[str, Any],
+        image_dirs: Dict[str, Path],
+        seg_mask_dirs: Dict[str, Path],
+        flood_masks_dirs: Dict[str, Path],
+        image_ext: str = "",
+    ):
+        self.stage = stage
+        self.cfg = deepcopy(cfg)
+        self.data = data
+        self.image_dirs = image_dirs
+        self.seg_mask_dirs = seg_mask_dirs
+        self.flood_masks_dirs = flood_masks_dirs
+        self.names = names
+        self.image_ext = image_ext
+
+        tfs = []
+        self.tfs = tvf.Compose(tfs)
+        self.augmentations = self.get_augmentations()
+
+    def __len__(self):
+        return len(self.names)
+
+    def __getitem__(self, idx):
+        if self.stage == "train" and self.cfg.random:
+            seed = None
+        else:
+            seed = [self.cfg.seed, idx]
+        (seed,) = np.random.SeedSequence(seed).generate_state(1)
+
+        scene, seq, name = self.names[idx]
+
+        view = self.get_view(
+            idx, scene, seq, name, seed
+        )
+
+        return view
+
+    def get_augmentations(self):
+        if self.stage != "train" or not self.cfg.augmentations.enabled:
+            print(f"No Augmentation!", "\n" * 10)
+            self.cfg.augmentations.random_flip = 0.0
+            return tvf.Compose([])
+
+        print(f"Augmentation!", "\n" * 10)
+        augmentations = [
+            tvf.ColorJitter(
+                brightness=self.cfg.augmentations.brightness,
+                contrast=self.cfg.augmentations.contrast,
+                saturation=self.cfg.augmentations.saturation,
+                hue=self.cfg.augmentations.hue,
+            )
+        ]
+
+        if self.cfg.augmentations.random_resized_crop:
+            augmentations.append(
+                tvf.RandomResizedCrop(scale=(0.8, 1.0))
+            )  # RandomResizedCrop
+
+        if self.cfg.augmentations.gaussian_noise.enabled:
+            augmentations.append(
+                tvf.GaussianNoise(
+                    mean=self.cfg.augmentations.gaussian_noise.mean,
+                    std=self.cfg.augmentations.gaussian_noise.std,
+                )
+            )  # Gaussian noise
+
+        if self.cfg.augmentations.brightness_contrast.enabled:
+            augmentations.append(
+                tvf.ColorJitter(
+                    brightness=self.cfg.augmentations.brightness_contrast.brightness_factor,
+                    contrast=self.cfg.augmentations.brightness_contrast.contrast_factor,
+                    saturation=0,  # Keep saturation at 0 for brightness and contrast adjustment
+                    hue=0,
+                )
+            )  # Brightness and contrast adjustment
+
+        return tvf.Compose(augmentations)
+
+    def random_flip(self, image, cam, valid, seg_mask, flood_mask, conf_mask):
+        if torch.rand(1) < self.cfg.augmentations.random_flip:
+            image = torch.flip(image, [-1])
+            cam = cam.flip()
+            valid = torch.flip(valid, [-1])
+            seg_mask = torch.flip(seg_mask, [1])
+            flood_mask = torch.flip(flood_mask, [-1])
+            conf_mask = torch.flip(conf_mask, [-1])
+
+        return image, cam, valid, seg_mask, flood_mask, conf_mask
+
+    def get_view(self, idx, scene, seq, name, seed):
+        data = {
+            "index": idx,
+            "name": name,
+            "scene": scene,
+            "sequence": seq,
+        }
+        cam_dict = self.data["cameras"][scene][seq][self.data["camera_id"][idx]]
+        cam = Camera.from_dict(cam_dict).float()
+
+        if "roll_pitch_yaw" in self.data:
+            roll, pitch, yaw = self.data["roll_pitch_yaw"][idx].numpy()
+        else:
+            roll, pitch, yaw = decompose_rotmat(
+                self.data["R_c2w"][idx].numpy())
+
+        image = read_image(self.image_dirs[scene] / (name + self.image_ext))
+        image = Image.fromarray(image)
+        image = self.augmentations(image)
+        image = np.array(image)
+
+        if "plane_params" in self.data:
+            # transform the plane parameters from world to camera frames
+            plane_w = self.data["plane_params"][idx]
+            data["ground_plane"] = torch.cat(
+                [rotmat2d(deg2rad(torch.tensor(yaw)))
+                 @ plane_w[:2], plane_w[2:]]
+            )
+
+        image, valid, cam, roll, pitch = self.process_image(
+            image, cam, roll, pitch, seed
+        )
+
+        if "chunk_index" in self.data:  # TODO: (cherie) do we need this?
+            data["chunk_id"] = (scene, seq, self.data["chunk_index"][idx])
+
+        # Semantic map extraction
+        seg_mask_path = self.seg_mask_dirs[scene] / \
+            (name.split("_")[0] + ".npy")
+        seg_masks_ours = np.load(seg_mask_path)
+        mask_center = (
+            seg_masks_ours.shape[0] // 2, seg_masks_ours.shape[1] // 2)
+
+        seg_masks_ours = seg_masks_ours[mask_center[0] -
+                                        100:mask_center[0], mask_center[1] - 50: mask_center[1] + 50]
+
+        if self.cfg.num_classes == 6:
+            seg_masks_ours = seg_masks_ours[..., [0, 1, 2, 4, 6, 7]]
+
+        flood_mask_path = self.flood_masks_dirs[scene] / \
+            (name.split("_")[0] + ".npy")
+        flood_mask = np.load(flood_mask_path)
+
+        flood_mask = flood_mask[mask_center[0]-100:mask_center[0],
+                                mask_center[1] - 50: mask_center[1] + 50]
+
+        confidence_map = flood_mask.copy()
+        confidence_map = (confidence_map - confidence_map.min()) / \
+            (confidence_map.max() - confidence_map.min() + 1e-6)
+
+        seg_masks_ours = torch.from_numpy(seg_masks_ours).float()
+        flood_mask = torch.from_numpy(flood_mask).float()
+        confidence_map = torch.from_numpy(confidence_map).float()
+
+        # Map Augmentations
+        with torch.random.fork_rng(devices=[]):
+            torch.manual_seed(seed)
+            image, cam, valid, seg_masks_ours, flood_mask, confidence_map = self.random_flip(
+                image, cam, valid, seg_masks_ours, flood_mask, confidence_map)
+
+        return {
+            **data,
+            "image": image,
+            "valid": valid,
+            "camera": cam,
+            "seg_masks": seg_masks_ours,
+            "flood_masks": flood_mask,
+            "roll_pitch_yaw": torch.tensor((roll, pitch, yaw)).float(),
+            "confidence_map": confidence_map
+            # "pixels_per_meter": torch.tensor(canvas.ppm).float(),
+        }
+
+    def process_image(self, image, cam, roll, pitch, seed):
+        image = (
+            torch.from_numpy(np.ascontiguousarray(image))
+            .permute(2, 0, 1)
+            .float()
+            .div_(255)
+        )
+
+        if not self.cfg.gravity_align:
+            # Turn off gravity alignment
+            roll = 0.0
+            pitch = 0.0
+            image, valid = rectify_image(image, cam, roll, pitch)
+        else:
+            image, valid = rectify_image(
+                image, cam, roll, pitch if self.cfg.rectify_pitch else None
+            )
+            roll = 0.0
+            if self.cfg.rectify_pitch:
+                pitch = 0.0
+
+        if self.cfg.target_focal_length is not None:
+            # Resize to a canonical focal length
+            factor = self.cfg.target_focal_length / cam.f.numpy()
+            size = (np.array(image.shape[-2:][::-1]) * factor).astype(int)
+            image, _, cam, valid = resize_image(
+                image, size, camera=cam, valid=valid)
+            size_out = self.cfg.resize_image
+            if size_out is None:
+                # Round the edges up such that they are multiple of a factor
+                stride = self.cfg.pad_to_multiple
+                size_out = (np.ceil((size / stride)) * stride).astype(int)
+            # Crop or pad such that both edges are of the given size
+            image, valid, cam = pad_image(
+                image, size_out, cam, valid, crop_and_center=True
+            )
+        elif self.cfg.resize_image is not None:
+            image, _, cam, valid = resize_image(
+                image, self.cfg.resize_image, fn=max, camera=cam, valid=valid
+            )
+            if self.cfg.pad_to_square:
+                # Pad such that both edges are of the given size
+                image, valid, cam = pad_image(
+                    image, self.cfg.resize_image, cam, valid)
+
+        if self.cfg.reduce_fov is not None:
+            h, w = image.shape[-2:]
+            f = float(cam.f[0])
+            fov = np.arctan(w / f / 2)
+            w_new = round(2 * f * np.tan(self.cfg.reduce_fov * fov))
+            image, valid, cam = pad_image(
+                image, (w_new, h), cam, valid, crop_and_center=True
+            )
+
+        with torch.random.fork_rng(devices=[]):
+            torch.manual_seed(seed)
+            image = self.tfs(image)
+
+        return image, valid, cam, roll, pitch

mapper/data/module.py

@@ -0,0 +1,64 @@
+from typing import Optional
+from omegaconf import DictConfig
+import pytorch_lightning as L
+import torch.utils.data as torchdata
+from .torch import collate, worker_init_fn
+
+
+def get_dataset(name):
+    if name == "mapillary":
+        from .mapillary.data_module import MapillaryDataModule
+        return MapillaryDataModule
+    elif name == "nuscenes":
+        from .nuscenes.data_module import NuScenesData
+        return NuScenesData
+    elif name == "kitti":
+        from .kitti.data_module import BEVKitti360Data
+        return BEVKitti360Data
+    else:
+        raise NotImplementedError(f"Dataset {name} not implemented.")
+
+
+class GenericDataModule(L.LightningDataModule):
+    def __init__(self, cfg: DictConfig):
+        super().__init__()
+        self.cfg = cfg
+        self.data_module = get_dataset(cfg.name)(cfg)
+
+    def prepare_data(self) -> None:
+        self.data_module.prepare_data()
+    
+    def setup(self, stage: Optional[str] = None):
+        self.data_module.setup(stage)
+
+    def dataloader(
+        self,
+        stage: str,
+        shuffle: bool = False,
+        num_workers: int = None,
+        sampler: Optional[torchdata.Sampler] = None,
+    ):
+        dataset = self.data_module.dataset(stage)
+        cfg = self.cfg["loading"][stage]
+        num_workers = cfg["num_workers"] if num_workers is None else num_workers
+        loader = torchdata.DataLoader(
+            dataset,
+            batch_size=cfg["batch_size"],
+            num_workers=num_workers,
+            shuffle=shuffle or (stage == "train"),
+            pin_memory=True,
+            persistent_workers=num_workers > 0,
+            worker_init_fn=worker_init_fn,
+            collate_fn=collate,
+            sampler=sampler,
+        )
+        return loader
+    
+    def train_dataloader(self, **kwargs):
+        return self.dataloader("train", **kwargs)
+
+    def val_dataloader(self, **kwargs):
+        return self.dataloader("val", **kwargs)
+
+    def test_dataloader(self, **kwargs):
+        return self.dataloader("test", **kwargs)

mapper/data/nuscenes/data_module.py

@@ -0,0 +1,33 @@
+from ..base import DataBase
+from .dataset import NuScenesDataset
+from ..schema import NuScenesDataConfiguration
+
+class NuScenesData(DataBase):
+    def __init__(self, cfg: NuScenesDataConfiguration):
+        self.cfg = cfg
+        self._dataset = {}
+
+    def prepare_data(self):
+        pass
+
+    def setup(self, stage):
+        if stage is None:
+            stage = 'fit'
+
+        split = {
+            'fit': 'train',
+            'val': 'val',
+            'validate': 'val',
+            'test': 'test'
+        }[stage]
+
+        self._dataset[split] = NuScenesDataset(
+            split=split,
+            cfg=self.cfg
+        )
+
+    def dataset(self, stage):
+        if self._dataset.get(stage) is None:
+            self.setup(stage)
+
+        return self._dataset[stage]

mapper/data/nuscenes/dataset.py

@@ -0,0 +1,207 @@
+import os
+import torch
+import numpy as np
+from pyquaternion import Quaternion
+from nuscenes.nuscenes import NuScenes
+from itertools import chain
+from PIL import Image
+from torchvision import transforms as T
+import torchvision.transforms as tvf
+from torchvision.transforms.functional import to_tensor
+
+from .splits_roddick import create_splits_scenes_roddick
+from ..image import pad_image, rectify_image, resize_image
+from .utils import decode_binary_labels
+from ..utils import decompose_rotmat
+from ...utils.io import read_image
+from ...utils.wrappers import Camera
+from ..schema import NuScenesDataConfiguration
+
+
+class NuScenesDataset(torch.utils.data.Dataset):
+    def __init__(self, cfg: NuScenesDataConfiguration, split="train"):
+
+        self.cfg = cfg
+        self.nusc = NuScenes(version=cfg.version, dataroot=str(cfg.data_dir))
+        self.map_data_root = cfg.map_dir
+        self.split = split
+
+        self.scenes = create_splits_scenes_roddick() # custom based on Roddick et al. 
+
+        scene_split = {
+            'v1.0-trainval': {'train': 'train', 'val': 'val', 'test': 'val'},
+            'v1.0-mini': {'train': 'mini_train', 'val': 'mini_val'},
+        }[cfg.version][split]
+        self.scenes = self.scenes[scene_split]
+        self.sample = list(filter(lambda sample: self.nusc.get(
+            'scene', sample['scene_token'])['name'] in self.scenes, self.nusc.sample))
+
+        self.tfs = self.get_augmentations() if split == "train" else T.Compose([])
+
+        data_tokens = []
+        for sample in self.sample:
+            data_token = sample['data']
+            data_token = [v for k,v in data_token.items() if k == "CAM_FRONT"]
+
+            data_tokens.append(data_token)
+
+        data_tokens = list(chain.from_iterable(data_tokens))
+        data = [self.nusc.get('sample_data', token) for token in data_tokens]
+
+        self.data = []
+        for d in data:
+            sample = self.nusc.get('sample', d['sample_token'])
+            scene = self.nusc.get('scene', sample['scene_token'])
+            location = self.nusc.get('log', scene['log_token'])['location']
+
+            file_name = d['filename']
+            ego_pose = self.nusc.get('ego_pose', d['ego_pose_token'])
+            calibrated_sensor = self.nusc.get(
+                "calibrated_sensor", d['calibrated_sensor_token'])
+
+            ego2global = np.eye(4).astype(np.float32)
+            ego2global[:3, :3] = Quaternion(ego_pose['rotation']).rotation_matrix
+            ego2global[:3, 3] = ego_pose['translation']
+
+            sensor2ego = np.eye(4).astype(np.float32)
+            sensor2ego[:3, :3] = Quaternion(
+                calibrated_sensor['rotation']).rotation_matrix
+            sensor2ego[:3, 3] = calibrated_sensor['translation']
+
+            sensor2global = ego2global @ sensor2ego
+
+            rotation = sensor2global[:3, :3]
+            roll, pitch, yaw = decompose_rotmat(rotation)
+
+            fx = calibrated_sensor['camera_intrinsic'][0][0]
+            fy = calibrated_sensor['camera_intrinsic'][1][1]
+            cx = calibrated_sensor['camera_intrinsic'][0][2]
+            cy = calibrated_sensor['camera_intrinsic'][1][2]
+            width = d['width']
+            height = d['height']
+
+            cam = Camera(torch.tensor(
+                [width, height, fx, fy, cx - 0.5, cy - 0.5])).float()
+            self.data.append({
+                'filename': file_name,
+                'yaw': yaw,
+                'pitch': pitch,
+                'roll': roll,
+                'cam': cam,
+                'sensor2global': sensor2global,
+                'token': d['token'],
+                'sample_token': d['sample_token'],
+                'location': location
+            })
+        
+        if self.cfg.percentage < 1.0 and split == "train":
+            self.data = self.data[:int(len(self.data) * self.cfg.percentage)]
+
+    def get_augmentations(self):
+
+        print(f"Augmentation!", "\n" * 10)
+        augmentations = [
+            tvf.ColorJitter(
+                brightness=self.cfg.augmentations.brightness,
+                contrast=self.cfg.augmentations.contrast,
+                saturation=self.cfg.augmentations.saturation,
+                hue=self.cfg.augmentations.hue,
+            )
+        ]
+
+        if self.cfg.augmentations.random_resized_crop:
+            augmentations.append(
+                tvf.RandomResizedCrop(scale=(0.8, 1.0))
+            )  # RandomResizedCrop
+
+        if self.cfg.augmentations.gaussian_noise.enabled:
+            augmentations.append(
+                tvf.GaussianNoise(
+                    mean=self.cfg.augmentations.gaussian_noise.mean,
+                    std=self.cfg.augmentations.gaussian_noise.std,
+                )
+            )  # Gaussian noise
+
+        if self.cfg.augmentations.brightness_contrast.enabled:
+            augmentations.append(
+                tvf.ColorJitter(
+                    brightness=self.cfg.augmentations.brightness_contrast.brightness_factor,
+                    contrast=self.cfg.augmentations.brightness_contrast.contrast_factor,
+                    saturation=0,  # Keep saturation at 0 for brightness and contrast adjustment
+                    hue=0,
+                )
+            )  # Brightness and contrast adjustment
+
+        return tvf.Compose(augmentations)
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, idx):
+        d = self.data[idx]
+
+        image = read_image(os.path.join(self.nusc.dataroot, d['filename']))
+        image = np.array(image)
+        cam = d['cam']
+        roll = d['roll']
+        pitch = d['pitch']
+        yaw = d['yaw']
+
+        with Image.open(self.map_data_root / f"{d['token']}.png") as semantic_image:
+            semantic_mask = to_tensor(semantic_image)
+
+        semantic_mask = decode_binary_labels(semantic_mask, self.cfg.num_classes + 1)
+        semantic_mask = torch.nn.functional.max_pool2d(semantic_mask.float(), (2, 2), stride=2) # 2 times downsample
+        semantic_mask = semantic_mask.permute(1, 2, 0)
+        semantic_mask = torch.flip(semantic_mask, [0])
+        
+        visibility_mask = semantic_mask[..., -1]
+        semantic_mask = semantic_mask[..., :-1]
+
+        if self.cfg.class_mapping is not None:
+            semantic_mask = semantic_mask[..., self.cfg.class_mapping]
+
+        image = (
+            torch.from_numpy(np.ascontiguousarray(image))
+            .permute(2, 0, 1)
+            .float()
+            .div_(255)
+        )
+
+        if not self.cfg.gravity_align:
+            # Turn off gravity alignment
+            roll = 0.0
+            pitch = 0.0
+            image, valid = rectify_image(image, cam, roll, pitch)
+
+        else:
+            image, valid = rectify_image(
+                image, cam, roll, pitch if self.cfg.rectify_pitch else None
+            )
+            roll = 0.0
+            if self.cfg.rectify_pitch:
+                pitch = 0.0
+        if self.cfg.resize_image is not None:
+            image, _, cam, valid = resize_image(
+                image, self.cfg.resize_image, fn=max, camera=cam, valid=valid
+            )
+            if self.cfg.pad_to_square:
+                image, valid, cam = pad_image(image, self.cfg.resize_image, cam, valid)
+        image = self.tfs(image)
+
+        confidence_map = visibility_mask.clone().float()
+        confidence_map = (confidence_map - confidence_map.min()) / (confidence_map.max() - confidence_map.min())
+
+        return {
+            "image": image,
+            "roll_pitch_yaw": torch.tensor([roll, pitch, yaw]).float(),
+            "camera": cam,
+            "valid": valid,
+            "seg_masks": semantic_mask.float(),
+            "token": d['token'],
+            "sample_token": d['sample_token'],
+            'location': d['location'],
+            'flood_masks': visibility_mask.float(),
+            "confidence_map": confidence_map,
+            'name': d['sample_token']
+        }

mapper/data/nuscenes/splits_roddick.py

@@ -0,0 +1,197 @@
+def create_splits_scenes_roddick():
+    train_roddick_scenes = [
+        "scene-0002", "scene-0003", "scene-0004", "scene-0005", "scene-0006", 
+        "scene-0007", "scene-0008", "scene-0009", "scene-0012", "scene-0013", 
+        "scene-0014", "scene-0015", "scene-0016", "scene-0017", "scene-0018", 
+        "scene-0019", "scene-0021", "scene-0022", "scene-0023", "scene-0024", 
+        "scene-0025", "scene-0026", "scene-0027", "scene-0028", "scene-0029", 
+        "scene-0030", "scene-0031", "scene-0032", "scene-0033", "scene-0034", 
+        "scene-0035", "scene-0036", "scene-0039", "scene-0042", "scene-0043", 
+        "scene-0044", "scene-0045", "scene-0046", "scene-0047", "scene-0048", 
+        "scene-0049", "scene-0050", "scene-0051", "scene-0052", "scene-0055", 
+        "scene-0056", "scene-0057", "scene-0058", "scene-0059", "scene-0060", 
+        "scene-0061", "scene-0062", "scene-0063", "scene-0064", "scene-0065", 
+        "scene-0066", "scene-0067", "scene-0068", "scene-0069", "scene-0070", 
+        "scene-0071", "scene-0072", "scene-0073", "scene-0074", "scene-0075", 
+        "scene-0076", "scene-0092", "scene-0093", "scene-0094", "scene-0095", 
+        "scene-0096", "scene-0097", "scene-0098", "scene-0099", "scene-0100", 
+        "scene-0101", "scene-0102", "scene-0103", "scene-0104", "scene-0105", 
+        "scene-0106", "scene-0107", "scene-0108", "scene-0109", "scene-0110", 
+        "scene-0120", "scene-0123", "scene-0124", "scene-0125", "scene-0126", 
+        "scene-0127", "scene-0128", "scene-0129", "scene-0130", "scene-0131", 
+        "scene-0132", "scene-0133", "scene-0134", "scene-0135", "scene-0138", 
+        "scene-0149", "scene-0150", "scene-0151", "scene-0154", "scene-0155", 
+        "scene-0157", "scene-0158", "scene-0159", "scene-0161", "scene-0162", 
+        "scene-0163", "scene-0164", "scene-0165", "scene-0166", "scene-0167", 
+        "scene-0168", "scene-0170", "scene-0171", "scene-0172", "scene-0173", 
+        "scene-0174", "scene-0175", "scene-0176", "scene-0177", "scene-0178", 
+        "scene-0179", "scene-0180", "scene-0181", "scene-0182", "scene-0183", 
+        "scene-0185", "scene-0187", "scene-0188", "scene-0190", "scene-0191", 
+        "scene-0192", "scene-0193", "scene-0194", "scene-0195", "scene-0196", 
+        "scene-0199", "scene-0200", "scene-0202", "scene-0203", "scene-0204", 
+        "scene-0206", "scene-0207", "scene-0208", "scene-0209", "scene-0210", 
+        "scene-0211", "scene-0212", "scene-0213", "scene-0214", "scene-0218", 
+        "scene-0219", "scene-0220", "scene-0221", "scene-0222", "scene-0224", 
+        "scene-0225", "scene-0226", "scene-0227", "scene-0228", "scene-0229", 
+        "scene-0230", "scene-0231", "scene-0232", "scene-0233", "scene-0234", 
+        "scene-0235", "scene-0236", "scene-0237", "scene-0238", "scene-0239", 
+        "scene-0240", "scene-0241", "scene-0242", "scene-0243", "scene-0244", 
+        "scene-0245", "scene-0246", "scene-0247", "scene-0248", "scene-0249", 
+        "scene-0250", "scene-0251", "scene-0252", "scene-0253", "scene-0254", 
+        "scene-0255", "scene-0256", "scene-0257", "scene-0258", "scene-0259", 
+        "scene-0260", "scene-0261", "scene-0262", "scene-0263", "scene-0264", 
+        "scene-0268", "scene-0270", "scene-0271", "scene-0272", "scene-0273", 
+        "scene-0274", "scene-0275", "scene-0276", "scene-0277", "scene-0278", 
+        "scene-0283", "scene-0284", "scene-0285", "scene-0286", "scene-0287", 
+        "scene-0288", "scene-0289", "scene-0290", "scene-0291", "scene-0292", 
+        "scene-0293", "scene-0294", "scene-0295", "scene-0296", "scene-0297", 
+        "scene-0298", "scene-0299", "scene-0300", "scene-0301", "scene-0302", 
+        "scene-0303", "scene-0304", "scene-0305", "scene-0306", "scene-0315", 
+        "scene-0316", "scene-0317", "scene-0318", "scene-0321", "scene-0323", 
+        "scene-0324", "scene-0328", "scene-0329", "scene-0330", "scene-0331", 
+        "scene-0332", "scene-0344", "scene-0345", "scene-0346", "scene-0349", 
+        "scene-0350", "scene-0351", "scene-0352", "scene-0353", "scene-0354", 
+        "scene-0355", "scene-0356", "scene-0357", "scene-0358", "scene-0359", 
+        "scene-0360", "scene-0361", "scene-0362", "scene-0363", "scene-0364", 
+        "scene-0365", "scene-0367", "scene-0370", "scene-0371", "scene-0372", 
+        "scene-0373", "scene-0374", "scene-0375", "scene-0376", "scene-0377", 
+        "scene-0379", "scene-0380", "scene-0381", "scene-0382", "scene-0383", 
+        "scene-0384", "scene-0385", "scene-0386", "scene-0388", "scene-0399", 
+        "scene-0400", "scene-0401", "scene-0402", "scene-0403", "scene-0405", 
+        "scene-0406", "scene-0407", "scene-0408", "scene-0420", "scene-0421", 
+        "scene-0422", "scene-0423", "scene-0424", "scene-0425", "scene-0426", 
+        "scene-0427", "scene-0428", "scene-0429", "scene-0430", "scene-0431", 
+        "scene-0432", "scene-0433", "scene-0434", "scene-0435", "scene-0436", 
+        "scene-0437", "scene-0438", "scene-0439", "scene-0440", "scene-0441", 
+        "scene-0442", "scene-0443", "scene-0444", "scene-0445", "scene-0446", 
+        "scene-0447", "scene-0448", "scene-0449", "scene-0450", "scene-0451", 
+        "scene-0452", "scene-0453", "scene-0454", "scene-0455", "scene-0456", 
+        "scene-0457", "scene-0458", "scene-0459", "scene-0461", "scene-0462", 
+        "scene-0463", "scene-0464", "scene-0465", "scene-0467", "scene-0468", 
+        "scene-0469", "scene-0471", "scene-0472", "scene-0474", "scene-0475", 
+        "scene-0476", "scene-0477", "scene-0478", "scene-0479", "scene-0480", 
+        "scene-0499", "scene-0500", "scene-0501", "scene-0502", "scene-0504", 
+        "scene-0505", "scene-0506", "scene-0507", "scene-0508", "scene-0509", 
+        "scene-0510", "scene-0511", "scene-0512", "scene-0513", "scene-0514", 
+        "scene-0515", "scene-0517", "scene-0518", "scene-0519", "scene-0520", 
+        "scene-0521", "scene-0522", "scene-0523", "scene-0524", "scene-0552", 
+        "scene-0553", "scene-0554", "scene-0555", "scene-0559", "scene-0560", 
+        "scene-0561", "scene-0562", "scene-0563", "scene-0564", "scene-0565", 
+        "scene-0584", "scene-0585", "scene-0586", "scene-0587", "scene-0588", 
+        "scene-0589", "scene-0590", "scene-0591", "scene-0592", "scene-0593", 
+        "scene-0594", "scene-0595", "scene-0596", "scene-0597", "scene-0598", 
+        "scene-0599", "scene-0600", "scene-0625", "scene-0626", "scene-0627", 
+        "scene-0629", "scene-0630", "scene-0632", "scene-0633", "scene-0634", 
+        "scene-0635", "scene-0636", "scene-0637", "scene-0638", "scene-0639", 
+        "scene-0640", "scene-0652", "scene-0653", "scene-0654", "scene-0655", 
+        "scene-0656", "scene-0657", "scene-0658", "scene-0659", "scene-0660", 
+        "scene-0661", "scene-0662", "scene-0663", "scene-0664", "scene-0665", 
+        "scene-0666", "scene-0667", "scene-0668", "scene-0669", "scene-0670", 
+        "scene-0671", "scene-0672", "scene-0673", "scene-0674", "scene-0675", 
+        "scene-0676", "scene-0677", "scene-0678", "scene-0679", "scene-0681", 
+        "scene-0683", "scene-0684", "scene-0685", "scene-0686", "scene-0687", 
+        "scene-0688", "scene-0689", "scene-0695", "scene-0696", "scene-0697", 
+        "scene-0698", "scene-0700", "scene-0701", "scene-0703", "scene-0704", 
+        "scene-0705", "scene-0706", "scene-0707", "scene-0708", "scene-0709", 
+        "scene-0710", "scene-0711", "scene-0712", "scene-0713", "scene-0714", 
+        "scene-0715", "scene-0716", "scene-0717", "scene-0718", "scene-0719", 
+        "scene-0726", "scene-0727", "scene-0728", "scene-0730", "scene-0731", 
+        "scene-0733", "scene-0734", "scene-0735", "scene-0736", "scene-0737", 
+        "scene-0738", "scene-0780", "scene-0781", "scene-0782", "scene-0783", 
+        "scene-0784", "scene-0786", "scene-0787", "scene-0789", "scene-0790", 
+        "scene-0791", "scene-0792", "scene-0802", "scene-0806", "scene-0808", 
+        "scene-0809", "scene-0810", "scene-0811", "scene-0812", "scene-0813", 
+        "scene-0815", "scene-0816", "scene-0817", "scene-0819", "scene-0820", 
+        "scene-0821", "scene-0822", "scene-0847", "scene-0848", "scene-0849", 
+        "scene-0850", "scene-0851", "scene-0852", "scene-0853", "scene-0854", 
+        "scene-0855", "scene-0856", "scene-0858", "scene-0860", "scene-0861", 
+        "scene-0862", "scene-0863", "scene-0864", "scene-0865", "scene-0866", 
+        "scene-0868", "scene-0869", "scene-0870", "scene-0871", "scene-0872", 
+        "scene-0873", "scene-0875", "scene-0876", "scene-0877", "scene-0878", 
+        "scene-0880", "scene-0882", "scene-0883", "scene-0884", "scene-0885", 
+        "scene-0886", "scene-0887", "scene-0888", "scene-0889", "scene-0890", 
+        "scene-0891", "scene-0892", "scene-0893", "scene-0894", "scene-0895", 
+        "scene-0896", "scene-0897", "scene-0898", "scene-0899", "scene-0900", 
+        "scene-0901", "scene-0902", "scene-0903", "scene-0904", "scene-0905", 
+        "scene-0906", "scene-0907", "scene-0908", "scene-0909", "scene-0916", 
+        "scene-0917", "scene-0921", "scene-0922", "scene-0923", "scene-0925", 
+        "scene-0926", "scene-0927", "scene-0928", "scene-0929", "scene-0930", 
+        "scene-0931", "scene-0945", "scene-0947", "scene-0949", "scene-0952", 
+        "scene-0953", "scene-0955", "scene-0956", "scene-0957", "scene-0958", 
+        "scene-0959", "scene-0960", "scene-0961", "scene-0966", "scene-0967", 
+        "scene-0968", "scene-0969", "scene-0971", "scene-0972", "scene-0975", 
+        "scene-0976", "scene-0977", "scene-0978", "scene-0979", "scene-0980", 
+        "scene-0981", "scene-0982", "scene-0983", "scene-0984", "scene-0988", 
+        "scene-0989", "scene-0990", "scene-0991", "scene-0992", "scene-0994", 
+        "scene-0995", "scene-0996", "scene-0997", "scene-0998", "scene-0999", 
+        "scene-1000", "scene-1001", "scene-1004", "scene-1005", "scene-1006", 
+        "scene-1007", "scene-1008", "scene-1009", "scene-1010", "scene-1011", 
+        "scene-1012", "scene-1013", "scene-1014", "scene-1015", "scene-1019", 
+        "scene-1020", "scene-1021", "scene-1022", "scene-1023", "scene-1024", 
+        "scene-1025", "scene-1044", "scene-1045", "scene-1046", "scene-1047", 
+        "scene-1048", "scene-1049", "scene-1050", "scene-1051", "scene-1052", 
+        "scene-1053", "scene-1054", "scene-1064", "scene-1065", "scene-1066", 
+        "scene-1067", "scene-1068", "scene-1069", "scene-1070", "scene-1071", 
+        "scene-1072", "scene-1073", "scene-1074", "scene-1075", "scene-1076", 
+        "scene-1077", "scene-1078", "scene-1079", "scene-1080", "scene-1081", 
+        "scene-1082", "scene-1083", "scene-1084", "scene-1085", "scene-1086", 
+        "scene-1087", "scene-1088", "scene-1089", "scene-1090", "scene-1091", 
+        "scene-1092", "scene-1093", "scene-1094", "scene-1095", "scene-1096", 
+        "scene-1097", "scene-1098", "scene-1099", "scene-1100", "scene-1101", 
+        "scene-1102", "scene-1104", "scene-1105", "scene-1106", "scene-1107", 
+        "scene-1108", "scene-1109", "scene-1110"]
+
+    val_roddick_scenes = [
+        "scene-0001", "scene-0010", "scene-0011", "scene-0020", "scene-0038", 
+        "scene-0041", "scene-0053", "scene-0054", "scene-0121", "scene-0122", 
+        "scene-0139", "scene-0152", "scene-0160", "scene-0184", "scene-0269", 
+        "scene-0347", "scene-0348", "scene-0366", "scene-0368", "scene-0369", 
+        "scene-0378", "scene-0389", "scene-0390", "scene-0391", "scene-0392", 
+        "scene-0393", "scene-0394", "scene-0395", "scene-0396", "scene-0397", 
+        "scene-0398", "scene-0411", "scene-0412", "scene-0413", "scene-0414", 
+        "scene-0415", "scene-0416", "scene-0417", "scene-0418", "scene-0419", 
+        "scene-0525", "scene-0526", "scene-0527", "scene-0528", "scene-0529", 
+        "scene-0530", "scene-0531", "scene-0532", "scene-0533", "scene-0534", 
+        "scene-0535", "scene-0536", "scene-0537", "scene-0538", "scene-0539", 
+        "scene-0541", "scene-0542", "scene-0543", "scene-0544", "scene-0545", 
+        "scene-0546", "scene-0556", "scene-0557", "scene-0558", "scene-0566", 
+        "scene-0568", "scene-0570", "scene-0571", "scene-0572", "scene-0573", 
+        "scene-0574", "scene-0575", "scene-0576", "scene-0577", "scene-0578", 
+        "scene-0580", "scene-0582", "scene-0583", "scene-0642", "scene-0643", 
+        "scene-0644", "scene-0645", "scene-0646", "scene-0647", "scene-0648", 
+        "scene-0649", "scene-0650", "scene-0651", "scene-0739", "scene-0740", 
+        "scene-0741", "scene-0744", "scene-0746", "scene-0747", "scene-0749", 
+        "scene-0750", "scene-0751", "scene-0752", "scene-0757", "scene-0758", 
+        "scene-0759", "scene-0760", "scene-0761", "scene-0762", "scene-0763", 
+        "scene-0764", "scene-0765", "scene-0767", "scene-0768", "scene-0769", 
+        "scene-0770", "scene-0771", "scene-0775", "scene-0777", "scene-0778", 
+        "scene-0794", "scene-0795", "scene-0796", "scene-0797", "scene-0798", 
+        "scene-0799", "scene-0800", "scene-0803", "scene-0804", "scene-0911", 
+        "scene-0912", "scene-0913", "scene-0914", "scene-0915", "scene-0919", 
+        "scene-0920", "scene-0924", "scene-0962", "scene-0963", "scene-1002", 
+        "scene-1003", "scene-1016", "scene-1017", "scene-1018", "scene-1055", 
+        "scene-1056", "scene-1057", "scene-1058", "scene-1059", "scene-1060", 
+        "scene-1061", "scene-1062", "scene-1063"]
+
+
+    calibration_roddick_scenes = [
+        "scene-0852", "scene-0429", "scene-0956", "scene-0194", "scene-0811", 
+        "scene-1110", "scene-1107", "scene-0294", "scene-0900", "scene-0596", 
+        "scene-0296", "scene-0885", "scene-0866", "scene-0105", "scene-0782", 
+        "scene-0191", "scene-0876", "scene-0133", "scene-0231", "scene-0847", 
+        "scene-0363", "scene-0026", "scene-0791", "scene-0909", "scene-0002", 
+        "scene-0283", "scene-0007", "scene-0251", "scene-1100", "scene-0668", 
+        "scene-0584", "scene-0287", "scene-0260", "scene-0171", "scene-0789", 
+        "scene-0108", "scene-0190", "scene-0206", "scene-0635", "scene-0815", 
+        "scene-0058", "scene-0710", "scene-0302", "scene-0639", "scene-0166", 
+        "scene-0094", "scene-0735", "scene-0321", "scene-1091", "scene-0344"
+    ]
+
+
+    scenes_dict = {
+        "train": train_roddick_scenes,
+        "val": val_roddick_scenes,
+        "calibration": calibration_roddick_scenes
+    }
+
+    return scenes_dict

mapper/data/nuscenes/utils.py

@@ -0,0 +1,214 @@
+import os
+import numpy as np
+from shapely import geometry, affinity
+from pyquaternion import Quaternion
+import cv2
+
+from nuscenes.eval.detection.utils import category_to_detection_name
+from nuscenes.eval.detection.constants import DETECTION_NAMES
+from nuscenes.utils.data_classes import LidarPointCloud
+
+from nuscenes.map_expansion.map_api import NuScenesMap
+from shapely.strtree import STRtree
+from collections import OrderedDict
+import torch
+
+def decode_binary_labels(labels, nclass):
+    bits = torch.pow(2, torch.arange(nclass))
+    return (labels & bits.view(-1, 1, 1)) > 0
+
+def transform_polygon(polygon, affine):
+    """
+    Transform a 2D polygon
+    """
+    a, b, tx, c, d, ty = affine.flatten()[:6]
+    return affinity.affine_transform(polygon, [a, b, c, d, tx, ty])
+
+
+def render_polygon(mask, polygon, extents, resolution, value=1):
+    if len(polygon) == 0:
+        return
+    polygon = (polygon - np.array(extents[:2])) / resolution
+    polygon = np.ascontiguousarray(polygon).round().astype(np.int32)
+    cv2.fillConvexPoly(mask, polygon, value)
+
+def transform(matrix, vectors):
+    vectors = np.dot(matrix[:-1, :-1], vectors.T)
+    vectors = vectors.T + matrix[:-1, -1]
+    return vectors
+
+CAMERA_NAMES = ['CAM_FRONT', 'CAM_FRONT_LEFT', 'CAM_FRONT_RIGHT', 
+                'CAM_BACK_LEFT', 'CAM_BACK_RIGHT', 'CAM_BACK']
+
+NUSCENES_CLASS_NAMES = [
+    'drivable_area', 'ped_crossing', 'walkway', 'carpark', 'car', 'truck', 
+    'bus', 'trailer', 'construction_vehicle', 'pedestrian', 'motorcycle', 
+    'bicycle', 'traffic_cone', 'barrier'
+]
+
+STATIC_CLASSES = ['drivable_area', 'ped_crossing', 'walkway', 'carpark_area']
+
+LOCATIONS = ['boston-seaport', 'singapore-onenorth', 'singapore-queenstown',
+             'singapore-hollandvillage']
+
+def load_map_data(dataroot, location):
+
+    # Load the NuScenes map object
+    nusc_map = NuScenesMap(dataroot, location)
+
+    map_data = OrderedDict()
+    for layer in STATIC_CLASSES:
+        
+        # Retrieve all data associated with the current layer
+        records = getattr(nusc_map, layer)
+        polygons = list()
+
+        # Drivable area records can contain multiple polygons
+        if layer == 'drivable_area':
+            for record in records:
+
+                # Convert each entry in the record into a shapely object
+                for token in record['polygon_tokens']:
+                    poly = nusc_map.extract_polygon(token)
+                    if poly.is_valid:
+                        polygons.append(poly)
+        else:
+            for record in records:
+
+                # Convert each entry in the record into a shapely object
+                poly = nusc_map.extract_polygon(record['polygon_token'])
+                if poly.is_valid:
+                    polygons.append(poly)
+
+        
+        # Store as an R-Tree for fast intersection queries
+        map_data[layer] = STRtree(polygons)
+    
+    return map_data
+
+def iterate_samples(nuscenes, start_token):
+    sample_token = start_token
+    while sample_token != '':
+        sample = nuscenes.get('sample', sample_token)
+        yield sample
+        sample_token = sample['next']
+    
+
+def get_map_masks(nuscenes, map_data, sample_data, extents, resolution):
+
+    # Render each layer sequentially
+    layers = [get_layer_mask(nuscenes, polys, sample_data, extents, 
+              resolution) for layer, polys in map_data.items()]
+
+    return np.stack(layers, axis=0)
+
+
+def get_layer_mask(nuscenes, polygons, sample_data, extents, resolution):
+
+    # Get the 2D affine transform from bev coords to map coords
+    tfm = get_sensor_transform(nuscenes, sample_data)[[0, 1, 3]][:, [0, 2, 3]]
+    inv_tfm = np.linalg.inv(tfm)
+
+    # Create a patch representing the birds-eye-view region in map coordinates
+    map_patch = geometry.box(*extents)
+    map_patch = transform_polygon(map_patch, tfm)
+
+    # Initialise the map mask
+    x1, z1, x2, z2 = extents
+    mask = np.zeros((int((z2 - z1) / resolution), int((x2 - x1) / resolution)),
+                    dtype=np.uint8)
+
+    # Find all polygons which intersect with the area of interest
+    for polygon in polygons.query(map_patch):
+
+        polygon = polygon.intersection(map_patch)
+        
+        # Transform into map coordinates
+        polygon = transform_polygon(polygon, inv_tfm)
+
+        # Render the polygon to the mask
+        render_shapely_polygon(mask, polygon, extents, resolution)
+    
+    return mask
+
+
+
+
+def get_object_masks(nuscenes, sample_data, extents, resolution):
+
+    # Initialize object masks
+    nclass = len(DETECTION_NAMES) + 1
+    grid_width = int((extents[2] - extents[0]) / resolution)
+    grid_height = int((extents[3] - extents[1]) / resolution)
+    masks = np.zeros((nclass, grid_height, grid_width), dtype=np.uint8)
+
+    # Get the 2D affine transform from bev coords to map coords
+    tfm = get_sensor_transform(nuscenes, sample_data)[[0, 1, 3]][:, [0, 2, 3]]
+    inv_tfm = np.linalg.inv(tfm)
+
+    for box in nuscenes.get_boxes(sample_data['token']):
+
+        # Get the index of the class
+        det_name = category_to_detection_name(box.name)
+        if det_name not in DETECTION_NAMES:
+            class_id = -1
+        else:
+            class_id = DETECTION_NAMES.index(det_name)
+        
+        # Get bounding box coordinates in the grid coordinate frame
+        bbox = box.bottom_corners()[:2]
+        local_bbox = np.dot(inv_tfm[:2, :2], bbox).T + inv_tfm[:2, 2]
+
+        # Render the rotated bounding box to the mask
+        render_polygon(masks[class_id], local_bbox, extents, resolution)
+    
+    return masks.astype(np.bool)
+
+
+def get_sensor_transform(nuscenes, sample_data):
+
+    # Load sensor transform data
+    sensor = nuscenes.get(
+        'calibrated_sensor', sample_data['calibrated_sensor_token'])
+    sensor_tfm = make_transform_matrix(sensor)
+
+    # Load ego pose data
+    pose = nuscenes.get('ego_pose', sample_data['ego_pose_token'])
+    pose_tfm = make_transform_matrix(pose)
+
+    return np.dot(pose_tfm, sensor_tfm)
+
+
+def load_point_cloud(nuscenes, sample_data):
+
+    # Load point cloud
+    lidar_path = os.path.join(nuscenes.dataroot, sample_data['filename'])
+    pcl = LidarPointCloud.from_file(lidar_path)
+    return pcl.points[:3, :].T
+
+
+def make_transform_matrix(record):
+    """
+    Create a 4x4 transform matrix from a calibrated_sensor or ego_pose record
+    """
+    transform = np.eye(4)
+    transform[:3, :3] = Quaternion(record['rotation']).rotation_matrix
+    transform[:3, 3] = np.array(record['translation'])
+    return transform
+
+
+def render_shapely_polygon(mask, polygon, extents, resolution):
+
+    if polygon.geom_type == 'Polygon':
+
+        # Render exteriors
+        render_polygon(mask, polygon.exterior.coords, extents, resolution, 1)
+
+        # Render interiors
+        for hole in polygon.interiors:
+            render_polygon(mask, hole.coords, extents, resolution, 0)
+    
+    # Handle the case of compound shapes
+    else:
+        for poly in polygon:
+            render_shapely_polygon(mask, poly, extents, resolution)

mapper/data/schema.py

@@ -0,0 +1,75 @@
+from dataclasses import dataclass
+from typing import Optional, Any, Dict
+from pathlib import Path
+
+@dataclass
+class AugmentationConfiguration:
+    gaussian_noise: dict
+    brightness_contrast: dict
+
+    enabled: bool = False
+    brightness: float = 0.5
+    contrast: float = 0.5
+    saturation: float = 0.5
+    hue: float = 0.5
+    random_resized_crop: Any = False
+    random_flip: float = 0.5
+
+
+@dataclass(kw_only=True)
+class DataConfiguration:
+    augmentations: AugmentationConfiguration
+
+    loading: Dict[str, Dict[str, Any]]
+
+    target_focal_length: Optional[int] = None
+    reduce_fov: Optional[bool] = None
+    resize_image: Optional[Any] = None
+    pad_to_square: Optional[bool] = None
+    pad_to_multiple: Optional[int] = None
+    gravity_align: Optional[bool] = None
+    rectify_pitch: Optional[bool] = True
+    num_classes: int
+
+    name: str
+    seed: Optional[int] = 0
+    random: Optional[bool] = True
+    num_threads: Optional[int] = None
+
+@dataclass(kw_only=True)
+class MIADataConfiguration(DataConfiguration):
+
+    scenes: list[str]
+    split: Any
+    data_dir: Path
+    pixel_per_meter: int
+    crop_size_meters: int
+
+    name: str = "mapillary"
+    filter_for: Optional[str] = None
+    filter_by_ground_angle: Optional[float] = None
+    min_num_points: int = 0
+
+@dataclass(kw_only=True)
+class KITTIDataConfiguration(DataConfiguration):
+    seam_root_dir: Path
+    dataset_root_dir: Path
+    bev_percentage: float
+
+    pixel_per_meter: int
+    crop_size_meters: int
+
+    class_mapping: Optional[Any] = None
+    percentage: float = 1.0
+
+@dataclass(kw_only=True)
+class NuScenesDataConfiguration(DataConfiguration):
+    data_dir: Path
+    map_dir: Path
+    pixel_per_meter: int
+    crop_size_meters: int
+
+    percentage: float = 1.0
+    class_mapping: Optional[Any] = None
+    version: str = "v1.0-trainval"
+    

mapper/data/sequential.py

@@ -0,0 +1,45 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+import numpy as np
+import torch
+
+
+def chunk_sequence(
+    data,
+    indices,
+    *,
+    names=None,
+    max_length=100,
+    min_length=1,
+    max_delay_s=None,
+    max_inter_dist=None,
+    max_total_dist=None,
+):
+    sort_array = data.get("capture_time", data.get("index"))
+    if sort_array is None:
+        sort_array = indices if names is None else names
+    indices = sorted(indices, key=lambda i: sort_array[i].tolist())
+    centers = torch.stack([data["t_c2w"][i][:2] for i in indices]).numpy()
+    dists = np.linalg.norm(np.diff(centers, axis=0), axis=-1)
+    if "capture_time" in data:
+        times = torch.stack([data["capture_time"][i] for i in indices])
+        times = times.double() / 1e3  # ms to s
+        delays = np.diff(times, axis=0)
+    else:
+        delays = np.zeros_like(dists)
+    chunks = [[indices[0]]]
+    dist_total = 0
+    for dist, delay, idx in zip(dists, delays, indices[1:]):
+        dist_total += dist
+        if (
+            (max_inter_dist is not None and dist > max_inter_dist)
+            or (max_total_dist is not None and dist_total > max_total_dist)
+            or (max_delay_s is not None and delay > max_delay_s)
+            or len(chunks[-1]) >= max_length
+        ):
+            chunks.append([])
+            dist_total = 0
+        chunks[-1].append(idx)
+    chunks = list(filter(lambda c: len(c) >= min_length, chunks))
+    chunks = sorted(chunks, key=len, reverse=True)
+    return chunks

mapper/data/torch.py

@@ -0,0 +1,102 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+import collections
+import os
+
+import torch
+from torch.utils.data import get_worker_info
+from torch.utils.data._utils.collate import (
+    default_collate_err_msg_format,
+    np_str_obj_array_pattern,
+)
+from lightning_fabric.utilities.seed import pl_worker_init_function
+
+def collate(batch):
+    """Difference with PyTorch default_collate: it can stack other tensor-like objects.
+    Adapted from PixLoc, Paul-Edouard Sarlin, ETH Zurich
+    https://github.com/cvg/pixloc
+    Released under the Apache License 2.0
+    """
+    if not isinstance(batch, list):  # no batching
+        return batch
+
+    # Filter None Elements
+    batch = [elem for elem in batch if elem is not None]
+    elem = batch[0]
+    elem_type = type(elem)
+    if isinstance(elem, torch.Tensor):
+        out = None
+        if torch.utils.data.get_worker_info() is not None:
+            # If we're in a background process, concatenate directly into a
+            # shared memory tensor to avoid an extra copy
+            numel = sum(x.numel() for x in batch)
+            storage = elem.storage()._new_shared(numel, device=elem.device)
+            out = elem.new(storage).resize_(len(batch), *list(elem.size()))
+        return torch.stack(batch, 0, out=out)
+    elif (
+        elem_type.__module__ == "numpy"
+        and elem_type.__name__ != "str_"
+        and elem_type.__name__ != "string_"
+    ):
+        if elem_type.__name__ == "ndarray" or elem_type.__name__ == "memmap":
+            # array of string classes and object
+            if np_str_obj_array_pattern.search(elem.dtype.str) is not None:
+                raise TypeError(default_collate_err_msg_format.format(elem.dtype))
+
+            return collate([torch.as_tensor(b) for b in batch])
+        elif elem.shape == ():  # scalars
+            return torch.as_tensor(batch)
+    elif isinstance(elem, float):
+        return torch.tensor(batch, dtype=torch.float64)
+    elif isinstance(elem, int):
+        return torch.tensor(batch)
+    elif isinstance(elem, (str, bytes)):
+        return batch
+    elif isinstance(elem, collections.abc.Mapping):
+        return {key: collate([d[key] for d in batch]) for key in elem}
+    elif isinstance(elem, tuple) and hasattr(elem, "_fields"):  # namedtuple
+        return elem_type(*(collate(samples) for samples in zip(*batch)))
+    elif isinstance(elem, collections.abc.Sequence):
+        # check to make sure that the elements in batch have consistent size
+        it = iter(batch)
+        elem_size = len(next(it))
+        if not all(len(elem) == elem_size for elem in it):
+            raise RuntimeError("each element in list of batch should be of equal size")
+        transposed = zip(*batch)
+        return [collate(samples) for samples in transposed]
+    else:
+        # try to stack anyway in case the object implements stacking.
+        try:
+            return torch.stack(batch, 0)
+        except TypeError as e:
+            if "expected Tensor as element" in str(e):
+                return batch
+            else:
+                raise e
+
+
+def set_num_threads(nt):
+    """Force numpy and other libraries to use a limited number of threads."""
+    try:
+        import mkl
+    except ImportError:
+        pass
+    else:
+        mkl.set_num_threads(nt)
+    torch.set_num_threads(1)
+    os.environ["IPC_ENABLE"] = "1"
+    for o in [
+        "OPENBLAS_NUM_THREADS",
+        "NUMEXPR_NUM_THREADS",
+        "OMP_NUM_THREADS",
+        "MKL_NUM_THREADS",
+    ]:
+        os.environ[o] = str(nt)
+
+
+def worker_init_fn(i):
+    info = get_worker_info()
+    pl_worker_init_function(info.id)
+    num_threads = info.dataset.cfg.get("num_threads")
+    if num_threads is not None:
+        set_num_threads(num_threads)

mapper/data/utils.py

@@ -0,0 +1,21 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+import numpy as np
+from scipy.spatial.transform import Rotation
+
+
+def crop_map(raster, xy, size, seed=None):
+    h, w = raster.shape[-2:]
+    state = np.random.RandomState(seed)
+    top = state.randint(0, h - size + 1)
+    left = state.randint(0, w - size + 1)
+    raster = raster[..., top : top + size, left : left + size]
+    xy -= np.array([left, top])
+    return raster, xy
+
+
+def decompose_rotmat(R_c2w):
+    R_cv2xyz = Rotation.from_euler("X", -90, degrees=True)
+    rot_w2c = R_cv2xyz * Rotation.from_matrix(R_c2w).inv()
+    roll, pitch, yaw = rot_w2c.as_euler("YXZ", degrees=True)
+    return roll, pitch, yaw

mapper/mapper.py

@@ -0,0 +1,112 @@
+import time
+import torch
+import hydra
+import pytorch_lightning as pl
+from typing import Any
+
+from hydra.core.config_store import ConfigStore
+from omegaconf import OmegaConf
+from pytorch_lightning.loggers import WandbLogger
+from pytorch_lightning.callbacks import ModelCheckpoint
+
+from pathlib import Path
+from dataclasses import dataclass
+
+from .module import GenericModule
+from .data.module import GenericDataModule
+from .callbacks import EvalSaveCallback, ImageLoggerCallback
+from .models.schema import ModelConfiguration, DINOConfiguration, ResNetConfiguration
+from .data.schema import MIADataConfiguration, KITTIDataConfiguration, NuScenesDataConfiguration
+
+
+@dataclass
+class ExperimentConfiguration:
+    name: str
+
+@dataclass
+class Configuration:
+    model: ModelConfiguration
+    experiment: ExperimentConfiguration
+    data: Any
+    training: Any
+
+
+cs = ConfigStore.instance()
+
+# Store root configuration schema
+cs.store(name="pretrain", node=Configuration)
+cs.store(name="mapper_nuscenes", node=Configuration)
+cs.store(name="mapper_kitti", node=Configuration)
+
+# Store data configuration schema
+cs.store(group="schema/data", name="mia",
+         node=MIADataConfiguration, package="data")
+cs.store(group="schema/data", name="kitti", node=KITTIDataConfiguration, package="data")
+cs.store(group="schema/data", name="nuscenes", node=NuScenesDataConfiguration, package="data")
+
+cs.store(group="model/schema/backbone", name="dino", node=DINOConfiguration, package="model.image_encoder.backbone")
+cs.store(group="model/schema/backbone", name="resnet", node=ResNetConfiguration, package="model.image_encoder.backbone")
+
+
+@hydra.main(version_base=None, config_path="conf", config_name="pretrain")
+def train(cfg: Configuration):
+    OmegaConf.resolve(cfg)
+
+    dm = GenericDataModule(cfg.data)
+
+    model = GenericModule(cfg)
+
+    exp_name_with_time = cfg.experiment.name + \
+        "_" + time.strftime("%Y-%m-%d_%H-%M-%S")
+
+    callbacks: list[pl.Callback]
+
+    if cfg.training.eval:
+        save_dir = Path(cfg.training.save_dir)
+        save_dir.mkdir(parents=True, exist_ok=True)
+
+        callbacks = [
+            EvalSaveCallback(save_dir=save_dir)
+        ]
+
+        logger = None
+    else:
+        callbacks = [
+            ImageLoggerCallback(num_classes=cfg.training.num_classes),
+            ModelCheckpoint(
+                monitor=cfg.training.checkpointing.monitor,
+                save_last=cfg.training.checkpointing.save_last,
+                save_top_k=cfg.training.checkpointing.save_top_k,
+            )
+        ]
+
+        logger = WandbLogger(
+            name=exp_name_with_time,
+            id=exp_name_with_time,
+            entity="mappred-large",
+            project="map-pred-full-v3",
+        )
+
+        logger.watch(model, log="all", log_freq=500)
+
+    if cfg.training.checkpoint is not None:
+        state_dict = torch.load(cfg.training.checkpoint)['state_dict']
+        model.load_state_dict(state_dict, strict=False)
+
+    trainer_args = OmegaConf.to_container(cfg.training.trainer)
+    trainer_args['callbacks'] = callbacks
+    trainer_args['logger'] = logger
+
+    trainer = pl.Trainer(**trainer_args)
+
+    if cfg.training.eval:
+        trainer.test(model, datamodule=dm)
+    else:
+        trainer.fit(model, datamodule=dm)
+
+
+if __name__ == "__main__":
+    pl.seed_everything(42)
+    torch.set_float32_matmul_precision("high")
+
+    train()

mapper/models/__init__.py

@@ -0,0 +1,28 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+# Adapted from PixLoc, Paul-Edouard Sarlin, ETH Zurich
+# https://github.com/cvg/pixloc
+# Released under the Apache License 2.0
+
+import inspect
+
+from .base import BaseModel
+
+
+def get_class(mod_name, base_path, BaseClass):
+    """Get the class object which inherits from BaseClass and is defined in
+    the module named mod_name, child of base_path.
+    """
+    mod_path = "{}.{}".format(base_path, mod_name)
+    mod = __import__(mod_path, fromlist=[""])
+    classes = inspect.getmembers(mod, inspect.isclass)
+    # Filter classes defined in the module
+    classes = [c for c in classes if c[1].__module__ == mod_path]
+    # Filter classes inherited from BaseModel
+    classes = [c for c in classes if issubclass(c[1], BaseClass)]
+    assert len(classes) == 1, classes
+    return classes[0][1]
+
+
+def get_model(name):
+    return get_class(name, __name__, BaseModel)

mapper/models/base.py

@@ -0,0 +1,59 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+# Adapted from PixLoc, Paul-Edouard Sarlin, ETH Zurich
+# https://github.com/cvg/pixloc
+# Released under the Apache License 2.0
+
+"""
+Base class for trainable models.
+"""
+
+from abc import ABCMeta, abstractmethod
+from copy import copy
+
+from omegaconf import OmegaConf
+from torch import nn
+
+
+class BaseModel(nn.Module, metaclass=ABCMeta):
+
+    required_data_keys = []
+    strict_conf = True
+
+    def __init__(self, conf):
+        """Perform some logic and call the _init method of the child model."""
+        super().__init__()
+        self.conf = conf 
+        OmegaConf.set_readonly(conf, True)
+        OmegaConf.set_struct(conf, True)
+        self.required_data_keys = copy(self.required_data_keys)
+        self._init(conf)
+
+    def forward(self, data):
+        """Check the data and call the _forward method of the child model."""
+
+        def recursive_key_check(expected, given):
+            for key in expected:
+                assert key in given, f"Missing key {key} in data"
+                if isinstance(expected, dict):
+                    recursive_key_check(expected[key], given[key])
+
+        recursive_key_check(self.required_data_keys, data)
+        return self._forward(data)
+
+    @abstractmethod
+    def _init(self, conf):
+        """To be implemented by the child class."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def _forward(self, data):
+        """To be implemented by the child class."""
+        raise NotImplementedError
+
+    def loss(self, pred, data):
+        """To be implemented by the child class."""
+        raise NotImplementedError
+
+    def metrics(self):
+        return {}  # no metrics

mapper/models/bev_projection.py

@@ -0,0 +1,95 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+
+import torch
+from torch.nn.functional import grid_sample
+
+from ..utils.geometry import from_homogeneous
+from .utils import make_grid
+
+
+class PolarProjectionDepth(torch.nn.Module):
+    def __init__(self, z_max, ppm, scale_range, z_min=None):
+        super().__init__()
+        self.z_max = z_max
+        self.Δ = Δ = 1 / ppm
+        self.z_min = z_min = Δ if z_min is None else z_min
+        self.scale_range = scale_range
+        z_steps = torch.arange(z_min, z_max + Δ, Δ)
+        self.register_buffer("depth_steps", z_steps, persistent=False)
+
+    def sample_depth_scores(self, pixel_scales, camera):
+        scale_steps = camera.f[..., None, 1] / self.depth_steps.flip(-1)
+        log_scale_steps = torch.log2(scale_steps)
+        scale_min, scale_max = self.scale_range
+        log_scale_norm = (log_scale_steps - scale_min) / \
+            (scale_max - scale_min)
+        log_scale_norm = log_scale_norm * 2 - 1  # in [-1, 1]
+
+        values = pixel_scales.flatten(1, 2).unsqueeze(-1)
+        indices = log_scale_norm.unsqueeze(-1)
+        indices = torch.stack([torch.zeros_like(indices), indices], -1)
+        depth_scores = grid_sample(values, indices, align_corners=True)
+        depth_scores = depth_scores.reshape(
+            pixel_scales.shape[:-1] + (len(self.depth_steps),)
+        )
+        return depth_scores
+
+    def forward(
+        self,
+        image,
+        pixel_scales,
+        camera,
+        return_total_score=False,
+    ):
+        depth_scores = self.sample_depth_scores(pixel_scales, camera)
+        depth_prob = torch.softmax(depth_scores, dim=1)
+        image_polar = torch.einsum("...dhw,...hwz->...dzw", image, depth_prob)
+        if return_total_score:
+            cell_score = torch.logsumexp(depth_scores, dim=1, keepdim=True)
+            return image_polar, cell_score.squeeze(1)
+        return image_polar
+
+
+class CartesianProjection(torch.nn.Module):
+    def __init__(self, z_max, x_max, ppm, z_min=None):
+        super().__init__()
+        self.z_max = z_max
+        self.x_max = x_max
+        self.Δ = Δ = 1 / ppm
+        self.z_min = z_min = Δ if z_min is None else z_min
+
+        grid_xz = make_grid(
+            x_max * 2 + Δ, z_max, step_y=Δ, step_x=Δ, orig_y=Δ, orig_x=-x_max, y_up=True
+        )
+        self.register_buffer("grid_xz", grid_xz, persistent=False)
+
+    def grid_to_polar(self, cam):
+        f, c = cam.f[..., 0][..., None, None], cam.c[..., 0][..., None, None]
+        u = from_homogeneous(self.grid_xz).squeeze(-1) * f + c
+        z_idx = (self.grid_xz[..., 1] - self.z_min) / \
+            self.Δ  # convert z value to index
+        z_idx = z_idx[None].expand_as(u)
+        grid_polar = torch.stack([u, z_idx], -1)
+        return grid_polar
+
+    def sample_from_polar(self, image_polar, valid_polar, grid_uz):
+        size = grid_uz.new_tensor(image_polar.shape[-2:][::-1])
+        grid_uz_norm = (grid_uz + 0.5) / size * 2 - 1
+        grid_uz_norm = grid_uz_norm * \
+            grid_uz.new_tensor([1, -1])  # y axis is up
+        image_bev = grid_sample(image_polar, grid_uz_norm, align_corners=False)
+
+        if valid_polar is None:
+            valid = torch.ones_like(image_polar[..., :1, :, :])
+        else:
+            valid = valid_polar.to(image_polar)[:, None]
+        valid = grid_sample(valid, grid_uz_norm, align_corners=False)
+        valid = valid.squeeze(1) > (1 - 1e-4)
+
+        return image_bev, valid
+
+    def forward(self, image_polar, valid_polar, cam):
+        grid_uz = self.grid_to_polar(cam)
+        image, valid = self.sample_from_polar(
+            image_polar, valid_polar, grid_uz)
+        return image, valid, grid_uz

mapper/models/dinov2/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+__version__ = "0.0.1"

mapper/models/dinov2/configs/__init__.py

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import pathlib
+
+from omegaconf import OmegaConf
+
+
+def load_config(config_name: str):
+    config_filename = config_name + ".yaml"
+    return OmegaConf.load(pathlib.Path(__file__).parent.resolve() / config_filename)
+
+
+dinov2_default_config = load_config("ssl_default_config")
+
+
+def load_and_merge_config(config_name: str):
+    default_config = OmegaConf.create(dinov2_default_config)
+    loaded_config = load_config(config_name)
+    return OmegaConf.merge(default_config, loaded_config)

mapper/models/dinov2/configs/eval/vitb14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_base
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

mapper/models/dinov2/configs/eval/vitg14_pretrain.yaml

@@ -0,0 +1,7 @@
+student:
+  arch: vit_giant2
+  patch_size: 14
+  ffn_layer: swiglufused
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

mapper/models/dinov2/configs/eval/vitl14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_large
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

mapper/models/dinov2/configs/eval/vits14_pretrain.yaml

@@ -0,0 +1,6 @@
+student:
+  arch: vit_small
+  patch_size: 14
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

mapper/models/dinov2/configs/eval/vits14_reg4_pretrain.yaml

@@ -0,0 +1,9 @@
+student:
+  arch: vit_small
+  patch_size: 14
+  num_register_tokens: 4
+  interpolate_antialias: true
+  interpolate_offset: 0.0
+crops:
+  global_crops_size: 518  # this is to set up the position embeddings properly
+  local_crops_size: 98

mapper/models/dinov2/configs/ssl_default_config.yaml

@@ -0,0 +1,118 @@
+MODEL:
+  WEIGHTS: ''
+compute_precision:
+  grad_scaler: true
+  teacher:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+  student:
+    backbone:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp16
+        buffer_dtype: fp32
+    dino_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+    ibot_head:
+      sharding_strategy: SHARD_GRAD_OP
+      mixed_precision:
+        param_dtype: fp16
+        reduce_dtype: fp32
+        buffer_dtype: fp32
+dino:
+  loss_weight: 1.0
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+  koleo_loss_weight: 0.1
+ibot:
+  loss_weight: 1.0
+  mask_sample_probability: 0.5
+  mask_ratio_min_max:
+  - 0.1
+  - 0.5
+  separate_head: false
+  head_n_prototypes: 65536
+  head_bottleneck_dim: 256
+  head_nlayers: 3
+  head_hidden_dim: 2048
+train:
+  batch_size_per_gpu: 64
+  dataset_path: ImageNet:split=TRAIN
+  output_dir: .
+  saveckp_freq: 20
+  seed: 0
+  num_workers: 10
+  OFFICIAL_EPOCH_LENGTH: 1250
+  cache_dataset: true
+  centering: "centering" # or "sinkhorn_knopp"
+student:
+  arch: vit_large
+  patch_size: 16
+  drop_path_rate: 0.3
+  layerscale: 1.0e-05
+  drop_path_uniform: true
+  pretrained_weights: ''
+  ffn_layer: "mlp"
+  block_chunks: 0
+  qkv_bias: true
+  proj_bias: true
+  ffn_bias: true
+  num_register_tokens: 0
+  interpolate_antialias: false
+  interpolate_offset: 0.1
+teacher:
+  momentum_teacher: 0.992
+  final_momentum_teacher: 1
+  warmup_teacher_temp: 0.04
+  teacher_temp: 0.07
+  warmup_teacher_temp_epochs: 30
+optim:
+  epochs: 100
+  weight_decay: 0.04
+  weight_decay_end: 0.4
+  base_lr: 0.004  # learning rate for a batch size of 1024
+  lr: 0.  # will be set after applying scaling rule
+  warmup_epochs: 10
+  min_lr: 1.0e-06
+  clip_grad: 3.0
+  freeze_last_layer_epochs: 1
+  scaling_rule: sqrt_wrt_1024
+  patch_embed_lr_mult: 0.2
+  layerwise_decay: 0.9
+  adamw_beta1: 0.9
+  adamw_beta2: 0.999
+crops:
+  global_crops_scale:
+  - 0.32
+  - 1.0
+  local_crops_number: 8
+  local_crops_scale:
+  - 0.05
+  - 0.32
+  global_crops_size: 224
+  local_crops_size: 96
+evaluation:
+  eval_period_iterations: 12500

mapper/models/dinov2/configs/train/vitg14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 12
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_giant2
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

mapper/models/dinov2/configs/train/vitl14.yaml

@@ -0,0 +1,26 @@
+dino:
+  head_n_prototypes: 131072
+  head_bottleneck_dim: 384
+ibot:
+  separate_head: true
+  head_n_prototypes: 131072
+train:
+  batch_size_per_gpu: 32
+  dataset_path: ImageNet22k
+  centering: sinkhorn_knopp
+student:
+  arch: vit_large
+  patch_size: 14
+  drop_path_rate: 0.4
+  ffn_layer: swiglufused
+  block_chunks: 4
+teacher:
+  momentum_teacher: 0.994
+optim:
+  epochs: 500
+  weight_decay_end: 0.2
+  base_lr: 2.0e-04  # learning rate for a batch size of 1024
+  warmup_epochs: 80
+  layerwise_decay: 1.0
+crops:
+  local_crops_size: 98

mapper/models/dinov2/configs/train/vitl16_short.yaml

@@ -0,0 +1,6 @@
+# this corresponds to the default config
+train:
+  dataset_path: ImageNet:split=TRAIN
+  batch_size_per_gpu: 64
+student:
+  block_chunks: 4

mapper/models/dinov2/data/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from .adapters import DatasetWithEnumeratedTargets
+from .loaders import make_data_loader, make_dataset, SamplerType
+from .collate import collate_data_and_cast
+from .masking import MaskingGenerator
+from .augmentations import DataAugmentationDINO

mapper/models/dinov2/data/adapters.py

@@ -0,0 +1,28 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+from typing import Any, Tuple
+
+from torch.utils.data import Dataset
+
+
+class DatasetWithEnumeratedTargets(Dataset):
+    def __init__(self, dataset):
+        self._dataset = dataset
+
+    def get_image_data(self, index: int) -> bytes:
+        return self._dataset.get_image_data(index)
+
+    def get_target(self, index: int) -> Tuple[Any, int]:
+        target = self._dataset.get_target(index)
+        return (index, target)
+
+    def __getitem__(self, index: int) -> Tuple[Any, Tuple[Any, int]]:
+        image, target = self._dataset[index]
+        target = index if target is None else target
+        return image, (index, target)
+
+    def __len__(self) -> int:
+        return len(self._dataset)

mapper/models/dinov2/data/augmentations.py

@@ -0,0 +1,118 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import logging
+
+from torchvision import transforms
+
+from .transforms import (
+    GaussianBlur,
+    make_normalize_transform,
+)
+
+
+logger = logging.getLogger("dinov2")
+
+
+class DataAugmentationDINO(object):
+    def __init__(
+        self,
+        global_crops_scale,
+        local_crops_scale,
+        local_crops_number,
+        global_crops_size=224,
+        local_crops_size=96,
+    ):
+        self.global_crops_scale = global_crops_scale
+        self.local_crops_scale = local_crops_scale
+        self.local_crops_number = local_crops_number
+        self.global_crops_size = global_crops_size
+        self.local_crops_size = local_crops_size
+
+        logger.info("###################################")
+        logger.info("Using data augmentation parameters:")
+        logger.info(f"global_crops_scale: {global_crops_scale}")
+        logger.info(f"local_crops_scale: {local_crops_scale}")
+        logger.info(f"local_crops_number: {local_crops_number}")
+        logger.info(f"global_crops_size: {global_crops_size}")
+        logger.info(f"local_crops_size: {local_crops_size}")
+        logger.info("###################################")
+
+        # random resized crop and flip
+        self.geometric_augmentation_global = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    global_crops_size, scale=global_crops_scale, interpolation=transforms.InterpolationMode.BICUBIC
+                ),
+                transforms.RandomHorizontalFlip(p=0.5),
+            ]
+        )
+
+        self.geometric_augmentation_local = transforms.Compose(
+            [
+                transforms.RandomResizedCrop(
+                    local_crops_size, scale=local_crops_scale, interpolation=transforms.InterpolationMode.BICUBIC
+                ),
+                transforms.RandomHorizontalFlip(p=0.5),
+            ]
+        )
+
+        # color distorsions / blurring
+        color_jittering = transforms.Compose(
+            [
+                transforms.RandomApply(
+                    [transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1)],
+                    p=0.8,
+                ),
+                transforms.RandomGrayscale(p=0.2),
+            ]
+        )
+
+        global_transfo1_extra = GaussianBlur(p=1.0)
+
+        global_transfo2_extra = transforms.Compose(
+            [
+                GaussianBlur(p=0.1),
+                transforms.RandomSolarize(threshold=128, p=0.2),
+            ]
+        )
+
+        local_transfo_extra = GaussianBlur(p=0.5)
+
+        # normalization
+        self.normalize = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                make_normalize_transform(),
+            ]
+        )
+
+        self.global_transfo1 = transforms.Compose([color_jittering, global_transfo1_extra, self.normalize])
+        self.global_transfo2 = transforms.Compose([color_jittering, global_transfo2_extra, self.normalize])
+        self.local_transfo = transforms.Compose([color_jittering, local_transfo_extra, self.normalize])
+
+    def __call__(self, image):
+        output = {}
+
+        # global crops:
+        im1_base = self.geometric_augmentation_global(image)
+        global_crop_1 = self.global_transfo1(im1_base)
+
+        im2_base = self.geometric_augmentation_global(image)
+        global_crop_2 = self.global_transfo2(im2_base)
+
+        output["global_crops"] = [global_crop_1, global_crop_2]
+
+        # global crops for teacher:
+        output["global_crops_teacher"] = [global_crop_1, global_crop_2]
+
+        # local crops:
+        local_crops = [
+            self.local_transfo(self.geometric_augmentation_local(image)) for _ in range(self.local_crops_number)
+        ]
+        output["local_crops"] = local_crops
+        output["offsets"] = ()
+
+        return output

mapper/models/dinov2/data/collate.py

@@ -0,0 +1,49 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the Apache License, Version 2.0
+# found in the LICENSE file in the root directory of this source tree.
+
+import torch
+import random
+
+
+def collate_data_and_cast(samples_list, mask_ratio_tuple, mask_probability, dtype, n_tokens=None, mask_generator=None):
+    # dtype = torch.half  # TODO: Remove
+
+    n_global_crops = len(samples_list[0][0]["global_crops"])
+    n_local_crops = len(samples_list[0][0]["local_crops"])
+
+    collated_global_crops = torch.stack([s[0]["global_crops"][i] for i in range(n_global_crops) for s in samples_list])
+
+    collated_local_crops = torch.stack([s[0]["local_crops"][i] for i in range(n_local_crops) for s in samples_list])
+
+    B = len(collated_global_crops)
+    N = n_tokens
+    n_samples_masked = int(B * mask_probability)
+    probs = torch.linspace(*mask_ratio_tuple, n_samples_masked + 1)
+    upperbound = 0
+    masks_list = []
+    for i in range(0, n_samples_masked):
+        prob_min = probs[i]
+        prob_max = probs[i + 1]
+        masks_list.append(torch.BoolTensor(mask_generator(int(N * random.uniform(prob_min, prob_max)))))
+        upperbound += int(N * prob_max)
+    for i in range(n_samples_masked, B):
+        masks_list.append(torch.BoolTensor(mask_generator(0)))
+
+    random.shuffle(masks_list)
+
+    collated_masks = torch.stack(masks_list).flatten(1)
+    mask_indices_list = collated_masks.flatten().nonzero().flatten()
+
+    masks_weight = (1 / collated_masks.sum(-1).clamp(min=1.0)).unsqueeze(-1).expand_as(collated_masks)[collated_masks]
+
+    return {
+        "collated_global_crops": collated_global_crops.to(dtype),
+        "collated_local_crops": collated_local_crops.to(dtype),
+        "collated_masks": collated_masks,
+        "mask_indices_list": mask_indices_list,
+        "masks_weight": masks_weight,
+        "upperbound": upperbound,
+        "n_masked_patches": torch.full((1,), fill_value=mask_indices_list.shape[0], dtype=torch.long),
+    }