init code

3rdparty/SCHP/__init__.py

@@ -0,0 +1,241 @@
+from collections import OrderedDict
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+from SCHP import networks
+from SCHP.utils.transforms import get_affine_transform, transform_logits
+from torchvision import transforms
+
+
+def get_palette(num_cls):
+    """Returns the color map for visualizing the segmentation mask.
+    Args:
+        num_cls: Number of classes
+    Returns:
+        The color map
+    """
+    n = num_cls
+    palette = [0] * (n * 3)
+    for j in range(0, n):
+        lab = j
+        palette[j * 3 + 0] = 0
+        palette[j * 3 + 1] = 0
+        palette[j * 3 + 2] = 0
+        i = 0
+        while lab:
+            palette[j * 3 + 0] |= ((lab >> 0) & 1) << (7 - i)
+            palette[j * 3 + 1] |= ((lab >> 1) & 1) << (7 - i)
+            palette[j * 3 + 2] |= ((lab >> 2) & 1) << (7 - i)
+            i += 1
+            lab >>= 3
+    return palette
+
+
+dataset_settings = {
+    "lip": {
+        "input_size": [473, 473],
+        "num_classes": 20,
+        "label": [
+            "Background",
+            "Hat",
+            "Hair",
+            "Glove",
+            "Sunglasses",
+            "Upper-clothes",
+            "Dress",
+            "Coat",
+            "Socks",
+            "Pants",
+            "Jumpsuits",
+            "Scarf",
+            "Skirt",
+            "Face",
+            "Left-arm",
+            "Right-arm",
+            "Left-leg",
+            "Right-leg",
+            "Left-shoe",
+            "Right-shoe",
+        ],
+    },
+    "atr": {
+        "input_size": [512, 512],
+        "num_classes": 18,
+        "label": [
+            "Background",
+            "Hat",
+            "Hair",
+            "Sunglasses",
+            "Upper-clothes",
+            "Skirt",
+            "Pants",
+            "Dress",
+            "Belt",
+            "Left-shoe",
+            "Right-shoe",
+            "Face",
+            "Left-leg",
+            "Right-leg",
+            "Left-arm",
+            "Right-arm",
+            "Bag",
+            "Scarf",
+        ],
+    },
+    "pascal": {
+        "input_size": [512, 512],
+        "num_classes": 7,
+        "label": [
+            "Background",
+            "Head",
+            "Torso",
+            "Upper Arms",
+            "Lower Arms",
+            "Upper Legs",
+            "Lower Legs",
+        ],
+    },
+}
+
+
+class SCHP:
+    def __init__(self, ckpt_path, device):
+        dataset_type = None
+        if "lip" in ckpt_path:
+            dataset_type = "lip"
+        elif "atr" in ckpt_path:
+            dataset_type = "atr"
+        elif "pascal" in ckpt_path:
+            dataset_type = "pascal"
+        assert dataset_type is not None, "Dataset type not found in checkpoint path"
+        self.device = device
+        self.num_classes = dataset_settings[dataset_type]["num_classes"]
+        self.input_size = dataset_settings[dataset_type]["input_size"]
+        self.aspect_ratio = self.input_size[1] * 1.0 / self.input_size[0]
+        self.palette = get_palette(self.num_classes)
+
+        self.label = dataset_settings[dataset_type]["label"]
+        self.model = networks.init_model(
+            "resnet101", num_classes=self.num_classes, pretrained=None
+        ).to(device)
+        self.load_ckpt(ckpt_path)
+        self.model.eval()
+
+        self.transform = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                transforms.Normalize(
+                    mean=[0.406, 0.456, 0.485], std=[0.225, 0.224, 0.229]
+                ),
+            ]
+        )
+        self.upsample = torch.nn.Upsample(
+            size=self.input_size, mode="bilinear", align_corners=True
+        )
+
+    def load_ckpt(self, ckpt_path):
+        rename_map = {
+            "decoder.conv3.2.weight": "decoder.conv3.3.weight",
+            "decoder.conv3.3.weight": "decoder.conv3.4.weight",
+            "decoder.conv3.3.bias": "decoder.conv3.4.bias",
+            "decoder.conv3.3.running_mean": "decoder.conv3.4.running_mean",
+            "decoder.conv3.3.running_var": "decoder.conv3.4.running_var",
+            "fushion.3.weight": "fushion.4.weight",
+            "fushion.3.bias": "fushion.4.bias",
+        }
+        state_dict = torch.load(ckpt_path, map_location="cpu")["state_dict"]
+        new_state_dict = OrderedDict()
+        for k, v in state_dict.items():
+            name = k[7:]  # remove `module.`
+            new_state_dict[name] = v
+        new_state_dict_ = OrderedDict()
+        for k, v in list(new_state_dict.items()):
+            if k in rename_map:
+                new_state_dict_[rename_map[k]] = v
+            else:
+                new_state_dict_[k] = v
+        self.model.load_state_dict(new_state_dict_, strict=False)
+
+    def _box2cs(self, box):
+        x, y, w, h = box[:4]
+        return self._xywh2cs(x, y, w, h)
+
+    def _xywh2cs(self, x, y, w, h):
+        center = np.zeros((2), dtype=np.float32)
+        center[0] = x + w * 0.5
+        center[1] = y + h * 0.5
+        if w > self.aspect_ratio * h:
+            h = w * 1.0 / self.aspect_ratio
+        elif w < self.aspect_ratio * h:
+            w = h * self.aspect_ratio
+        scale = np.array([w, h], dtype=np.float32)
+        return center, scale
+
+    def preprocess(self, image):
+        if isinstance(image, str):
+            img = cv2.imread(image, cv2.IMREAD_COLOR)
+        elif isinstance(image, Image.Image):
+            # to cv2 format
+            img = np.array(image)
+
+        h, w, _ = img.shape
+        # Get person center and scale
+        person_center, s = self._box2cs([0, 0, w - 1, h - 1])
+        r = 0
+        trans = get_affine_transform(person_center, s, r, self.input_size)
+        input = cv2.warpAffine(
+            img,
+            trans,
+            (int(self.input_size[1]), int(self.input_size[0])),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0, 0, 0),
+        )
+
+        input = self.transform(input).to(self.device).unsqueeze(0)
+        meta = {
+            "center": person_center,
+            "height": h,
+            "width": w,
+            "scale": s,
+            "rotation": r,
+        }
+        return input, meta
+
+    def __call__(self, image_or_path):
+        if isinstance(image_or_path, list):
+            image_list = []
+            meta_list = []
+            for image in image_or_path:
+                image, meta = self.preprocess(image)
+                image_list.append(image)
+                meta_list.append(meta)
+            image = torch.cat(image_list, dim=0)
+        else:
+            image, meta = self.preprocess(image_or_path)
+            meta_list = [meta]
+
+        output = self.model(image)
+        # upsample_outputs = self.upsample(output[0][-1])
+        upsample_outputs = self.upsample(output)
+        upsample_outputs = upsample_outputs.permute(0, 2, 3, 1)  # BCHW -> BHWC
+
+        output_img_list = []
+        for upsample_output, meta in zip(upsample_outputs, meta_list):
+            c, s, w, h = meta["center"], meta["scale"], meta["width"], meta["height"]
+            logits_result = transform_logits(
+                upsample_output.data.cpu().numpy(),
+                c,
+                s,
+                w,
+                h,
+                input_size=self.input_size,
+            )
+            parsing_result = np.argmax(logits_result, axis=2)
+            output_img = Image.fromarray(np.asarray(parsing_result, dtype=np.uint8))
+            output_img.putpalette(self.palette)
+            output_img_list.append(output_img)
+
+        return output_img_list[0] if len(output_img_list) == 1 else output_img_list

3rdparty/SCHP/networks/AugmentCE2P.py

@@ -0,0 +1,480 @@
+#!/usr/bin/env python
+# -*- encoding: utf-8 -*-
+
+"""
+@Author  :   Peike Li
+@Contact :   peike.li@yahoo.com
+@File    :   AugmentCE2P.py
+@Time    :   8/4/19 3:35 PM
+@Desc    :
+@License :   This source code is licensed under the license found in the
+             LICENSE file in the root directory of this source tree.
+"""
+
+import torch
+import torch.nn as nn
+
+from torch.nn import BatchNorm2d, functional as F, LeakyReLU
+
+affine_par = True
+pretrained_settings = {
+    "resnet101": {
+        "imagenet": {
+            "input_space": "BGR",
+            "input_size": [3, 224, 224],
+            "input_range": [0, 1],
+            "mean": [0.406, 0.456, 0.485],
+            "std": [0.225, 0.224, 0.229],
+            "num_classes": 1000,
+        }
+    },
+}
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    "3x3 convolution with padding"
+    return nn.Conv2d(
+        in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False
+    )
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(
+        self,
+        inplanes,
+        planes,
+        stride=1,
+        dilation=1,
+        downsample=None,
+        fist_dilation=1,
+        multi_grid=1,
+    ):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(
+            planes,
+            planes,
+            kernel_size=3,
+            stride=stride,
+            padding=dilation * multi_grid,
+            dilation=dilation * multi_grid,
+            bias=False,
+        )
+        self.bn2 = BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
+        self.bn3 = BatchNorm2d(planes * 4)
+        self.relu = nn.ReLU(inplace=False)
+        self.relu_inplace = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.dilation = dilation
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out = out + residual
+        out = self.relu_inplace(out)
+
+        return out
+
+
+class PSPModule(nn.Module):
+    """
+    Reference:
+        Zhao, Hengshuang, et al. *"Pyramid scene parsing network."*
+    """
+
+    def __init__(self, features, out_features=512, sizes=(1, 2, 3, 6)):
+        super(PSPModule, self).__init__()
+
+        self.stages = []
+        self.stages = nn.ModuleList(
+            [self._make_stage(features, out_features, size) for size in sizes]
+        )
+        self.bottleneck = nn.Sequential(
+            nn.Conv2d(
+                features + len(sizes) * out_features,
+                out_features,
+                kernel_size=3,
+                padding=1,
+                dilation=1,
+                bias=False,
+            ),
+            BatchNorm2d(out_features),
+            LeakyReLU(),
+        )
+
+    def _make_stage(self, features, out_features, size):
+        prior = nn.AdaptiveAvgPool2d(output_size=(size, size))
+        conv = nn.Conv2d(features, out_features, kernel_size=1, bias=False)
+        return nn.Sequential(
+            prior,
+            conv,
+            # bn
+            BatchNorm2d(out_features),
+            LeakyReLU(),
+        )
+
+    def forward(self, feats):
+        h, w = feats.size(2), feats.size(3)
+        priors = [
+            F.interpolate(
+                input=stage(feats), size=(h, w), mode="bilinear", align_corners=True
+            )
+            for stage in self.stages
+        ] + [feats]
+        bottle = self.bottleneck(torch.cat(priors, 1))
+        return bottle
+
+
+class ASPPModule(nn.Module):
+    """
+    Reference:
+        Chen, Liang-Chieh, et al. *"Rethinking Atrous Convolution for Semantic Image Segmentation."*
+    """
+
+    def __init__(
+        self, features, inner_features=256, out_features=512, dilations=(12, 24, 36)
+    ):
+        super(ASPPModule, self).__init__()
+
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, 1)),
+            nn.Conv2d(
+                features,
+                inner_features,
+                kernel_size=1,
+                padding=0,
+                dilation=1,
+                bias=False,
+            ),
+            #    InPlaceABNSync(inner_features)
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(
+                features,
+                inner_features,
+                kernel_size=1,
+                padding=0,
+                dilation=1,
+                bias=False,
+            ),
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+        )
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(
+                features,
+                inner_features,
+                kernel_size=3,
+                padding=dilations[0],
+                dilation=dilations[0],
+                bias=False,
+            ),
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+        )
+        self.conv4 = nn.Sequential(
+            nn.Conv2d(
+                features,
+                inner_features,
+                kernel_size=3,
+                padding=dilations[1],
+                dilation=dilations[1],
+                bias=False,
+            ),
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+        )
+        self.conv5 = nn.Sequential(
+            nn.Conv2d(
+                features,
+                inner_features,
+                kernel_size=3,
+                padding=dilations[2],
+                dilation=dilations[2],
+                bias=False,
+            ),
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+        )
+
+        self.bottleneck = nn.Sequential(
+            nn.Conv2d(
+                inner_features * 5,
+                out_features,
+                kernel_size=1,
+                padding=0,
+                dilation=1,
+                bias=False,
+            ),
+            BatchNorm2d(inner_features),
+            LeakyReLU(),
+            nn.Dropout2d(0.1),
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), 1)
+
+        bottle = self.bottleneck(out)
+        return bottle
+
+
+class Edge_Module(nn.Module):
+    """
+    Edge Learning Branch
+    """
+
+    def __init__(self, in_fea=[256, 512, 1024], mid_fea=256, out_fea=2):
+        super(Edge_Module, self).__init__()
+
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(
+                in_fea[0], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False
+            ),
+            BatchNorm2d(mid_fea),
+            LeakyReLU(),
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(
+                in_fea[1], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False
+            ),
+            BatchNorm2d(mid_fea),
+            LeakyReLU(),
+        )
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(
+                in_fea[2], mid_fea, kernel_size=1, padding=0, dilation=1, bias=False
+            ),
+            BatchNorm2d(mid_fea),
+            LeakyReLU(),
+        )
+        self.conv4 = nn.Conv2d(
+            mid_fea, out_fea, kernel_size=3, padding=1, dilation=1, bias=True
+        )
+        # self.conv5 = nn.Conv2d(out_fea * 3, out_fea, kernel_size=1, padding=0, dilation=1, bias=True)
+
+    def forward(self, x1, x2, x3):
+        _, _, h, w = x1.size()
+
+        edge1_fea = self.conv1(x1)
+        # edge1 = self.conv4(edge1_fea)
+        edge2_fea = self.conv2(x2)
+        edge2 = self.conv4(edge2_fea)
+        edge3_fea = self.conv3(x3)
+        edge3 = self.conv4(edge3_fea)
+
+        edge2_fea = F.interpolate(
+            edge2_fea, size=(h, w), mode="bilinear", align_corners=True
+        )
+        edge3_fea = F.interpolate(
+            edge3_fea, size=(h, w), mode="bilinear", align_corners=True
+        )
+        edge2 = F.interpolate(edge2, size=(h, w), mode="bilinear", align_corners=True)
+        edge3 = F.interpolate(edge3, size=(h, w), mode="bilinear", align_corners=True)
+
+        # edge = torch.cat([edge1, edge2, edge3], dim=1)
+        edge_fea = torch.cat([edge1_fea, edge2_fea, edge3_fea], dim=1)
+        # edge = self.conv5(edge)
+
+        # return edge, edge_fea
+        return edge_fea
+
+
+class Decoder_Module(nn.Module):
+    """
+    Parsing Branch Decoder Module.
+    """
+
+    def __init__(self, num_classes):
+        super(Decoder_Module, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(512, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            BatchNorm2d(256),
+            LeakyReLU(),
+        )
+        self.conv2 = nn.Sequential(
+            nn.Conv2d(
+                256, 48, kernel_size=1, stride=1, padding=0, dilation=1, bias=False
+            ),
+            BatchNorm2d(48),
+            LeakyReLU(),
+        )
+        self.conv3 = nn.Sequential(
+            nn.Conv2d(304, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            BatchNorm2d(256),
+            LeakyReLU(),
+            nn.Conv2d(256, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            BatchNorm2d(256),
+            LeakyReLU(),
+        )
+
+        # self.conv4 = nn.Conv2d(256, num_classes, kernel_size=1, padding=0, dilation=1, bias=True)
+
+    def forward(self, xt, xl):
+        _, _, h, w = xl.size()
+        xt = F.interpolate(
+            self.conv1(xt), size=(h, w), mode="bilinear", align_corners=True
+        )
+        xl = self.conv2(xl)
+        x = torch.cat([xt, xl], dim=1)
+        x = self.conv3(x)
+        # seg = self.conv4(x)
+        # return seg, x
+        return x
+
+
+class ResNet(nn.Module):
+    def __init__(self, block, layers, num_classes):
+        self.inplanes = 128
+        super(ResNet, self).__init__()
+        self.conv1 = conv3x3(3, 64, stride=2)
+        self.bn1 = BatchNorm2d(64)
+        self.relu1 = nn.ReLU(inplace=False)
+        self.conv2 = conv3x3(64, 64)
+        self.bn2 = BatchNorm2d(64)
+        self.relu2 = nn.ReLU(inplace=False)
+        self.conv3 = conv3x3(64, 128)
+        self.bn3 = BatchNorm2d(128)
+        self.relu3 = nn.ReLU(inplace=False)
+
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(
+            block, 512, layers[3], stride=1, dilation=2, multi_grid=(1, 1, 1)
+        )
+
+        self.context_encoding = PSPModule(2048, 512)
+
+        self.edge = Edge_Module()
+        self.decoder = Decoder_Module(num_classes)
+
+        self.fushion = nn.Sequential(
+            nn.Conv2d(1024, 256, kernel_size=1, padding=0, dilation=1, bias=False),
+            BatchNorm2d(256),
+            LeakyReLU(),
+            nn.Dropout2d(0.1),
+            nn.Conv2d(
+                256, num_classes, kernel_size=1, padding=0, dilation=1, bias=True
+            ),
+        )
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilation=1, multi_grid=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(
+                    self.inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    bias=False,
+                ),
+                BatchNorm2d(planes * block.expansion, affine=affine_par),
+            )
+
+        layers = []
+        generate_multi_grid = lambda index, grids: (
+            grids[index % len(grids)] if isinstance(grids, tuple) else 1
+        )
+        layers.append(
+            block(
+                self.inplanes,
+                planes,
+                stride,
+                dilation=dilation,
+                downsample=downsample,
+                multi_grid=generate_multi_grid(0, multi_grid),
+            )
+        )
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(
+                block(
+                    self.inplanes,
+                    planes,
+                    dilation=dilation,
+                    multi_grid=generate_multi_grid(i, multi_grid),
+                )
+            )
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.relu1(self.bn1(self.conv1(x)))
+        x = self.relu2(self.bn2(self.conv2(x)))
+        x = self.relu3(self.bn3(self.conv3(x)))
+        x = self.maxpool(x)
+        x2 = self.layer1(x)
+        x3 = self.layer2(x2)
+        x4 = self.layer3(x3)
+        x5 = self.layer4(x4)
+        x = self.context_encoding(x5)
+        # parsing_result, parsing_fea = self.decoder(x, x2)
+        parsing_fea = self.decoder(x, x2)
+        # Edge Branch
+        # edge_result, edge_fea = self.edge(x2, x3, x4)
+        edge_fea = self.edge(x2, x3, x4)
+        # Fusion Branch
+        x = torch.cat([parsing_fea, edge_fea], dim=1)
+        fusion_result = self.fushion(x)
+        # return [[parsing_result, fusion_result], [edge_result]]
+        return fusion_result
+
+
+def initialize_pretrained_model(
+    model, settings, pretrained="./models/resnet101-imagenet.pth"
+):
+    model.input_space = settings["input_space"]
+    model.input_size = settings["input_size"]
+    model.input_range = settings["input_range"]
+    model.mean = settings["mean"]
+    model.std = settings["std"]
+
+    if pretrained is not None:
+        saved_state_dict = torch.load(pretrained)
+        new_params = model.state_dict().copy()
+        for i in saved_state_dict:
+            i_parts = i.split(".")
+            if not i_parts[0] == "fc":
+                new_params[".".join(i_parts[0:])] = saved_state_dict[i]
+        model.load_state_dict(new_params)
+
+
+def resnet101(num_classes=20, pretrained="./models/resnet101-imagenet.pth"):
+    model = ResNet(Bottleneck, [3, 4, 23, 3], num_classes)
+    settings = pretrained_settings["resnet101"]["imagenet"]
+    initialize_pretrained_model(model, settings, pretrained)
+    return model

3rdparty/SCHP/networks/__init__.py

@@ -0,0 +1,13 @@
+from __future__ import absolute_import
+
+from SCHP.networks.AugmentCE2P import resnet101
+
+__factory = {
+    "resnet101": resnet101,
+}
+
+
+def init_model(name, *args, **kwargs):
+    if name not in __factory.keys():
+        raise KeyError("Unknown model arch: {}".format(name))
+    return __factory[name](*args, **kwargs)

3rdparty/SCHP/utils/transforms.py

@@ -0,0 +1,174 @@
+# ------------------------------------------------------------------------------
+# Copyright (c) Microsoft
+# Licensed under the MIT License.
+# Written by Bin Xiao (Bin.Xiao@microsoft.com)
+# ------------------------------------------------------------------------------
+
+from __future__ import absolute_import, division, print_function
+
+import cv2
+
+import numpy as np
+import torch
+
+
+class BRG2Tensor_transform(object):
+    def __call__(self, pic):
+        img = torch.from_numpy(pic.transpose((2, 0, 1)))
+        if isinstance(img, torch.ByteTensor):
+            return img.float()
+        else:
+            return img
+
+
+class BGR2RGB_transform(object):
+    def __call__(self, tensor):
+        return tensor[[2, 1, 0], :, :]
+
+
+def flip_back(output_flipped, matched_parts):
+    """
+    ouput_flipped: numpy.ndarray(batch_size, num_joints, height, width)
+    """
+    assert (
+        output_flipped.ndim == 4
+    ), "output_flipped should be [batch_size, num_joints, height, width]"
+
+    output_flipped = output_flipped[:, :, :, ::-1]
+
+    for pair in matched_parts:
+        tmp = output_flipped[:, pair[0], :, :].copy()
+        output_flipped[:, pair[0], :, :] = output_flipped[:, pair[1], :, :]
+        output_flipped[:, pair[1], :, :] = tmp
+
+    return output_flipped
+
+
+def fliplr_joints(joints, joints_vis, width, matched_parts):
+    """
+    flip coords
+    """
+    # Flip horizontal
+    joints[:, 0] = width - joints[:, 0] - 1
+
+    # Change left-right parts
+    for pair in matched_parts:
+        joints[pair[0], :], joints[pair[1], :] = (
+            joints[pair[1], :],
+            joints[pair[0], :].copy(),
+        )
+        joints_vis[pair[0], :], joints_vis[pair[1], :] = (
+            joints_vis[pair[1], :],
+            joints_vis[pair[0], :].copy(),
+        )
+
+    return joints * joints_vis, joints_vis
+
+
+def transform_preds(coords, center, scale, input_size):
+    target_coords = np.zeros(coords.shape)
+    trans = get_affine_transform(center, scale, 0, input_size, inv=1)
+    for p in range(coords.shape[0]):
+        target_coords[p, 0:2] = affine_transform(coords[p, 0:2], trans)
+    return target_coords
+
+
+def transform_parsing(pred, center, scale, width, height, input_size):
+
+    trans = get_affine_transform(center, scale, 0, input_size, inv=1)
+    target_pred = cv2.warpAffine(
+        pred,
+        trans,
+        (int(width), int(height)),  # (int(width), int(height)),
+        flags=cv2.INTER_NEAREST,
+        borderMode=cv2.BORDER_CONSTANT,
+        borderValue=(0),
+    )
+
+    return target_pred
+
+
+def transform_logits(logits, center, scale, width, height, input_size):
+
+    trans = get_affine_transform(center, scale, 0, input_size, inv=1)
+    channel = logits.shape[2]
+    target_logits = []
+    for i in range(channel):
+        target_logit = cv2.warpAffine(
+            logits[:, :, i],
+            trans,
+            (int(width), int(height)),  # (int(width), int(height)),
+            flags=cv2.INTER_LINEAR,
+            borderMode=cv2.BORDER_CONSTANT,
+            borderValue=(0),
+        )
+        target_logits.append(target_logit)
+    target_logits = np.stack(target_logits, axis=2)
+
+    return target_logits
+
+
+def get_affine_transform(
+    center, scale, rot, output_size, shift=np.array([0, 0], dtype=np.float32), inv=0
+):
+    if not isinstance(scale, np.ndarray) and not isinstance(scale, list):
+        print(scale)
+        scale = np.array([scale, scale])
+
+    scale_tmp = scale
+
+    src_w = scale_tmp[0]
+    dst_w = output_size[1]
+    dst_h = output_size[0]
+
+    rot_rad = np.pi * rot / 180
+    src_dir = get_dir([0, src_w * -0.5], rot_rad)
+    dst_dir = np.array([0, (dst_w - 1) * -0.5], np.float32)
+
+    src = np.zeros((3, 2), dtype=np.float32)
+    dst = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = center + scale_tmp * shift
+    src[1, :] = center + src_dir + scale_tmp * shift
+    dst[0, :] = [(dst_w - 1) * 0.5, (dst_h - 1) * 0.5]
+    dst[1, :] = np.array([(dst_w - 1) * 0.5, (dst_h - 1) * 0.5]) + dst_dir
+
+    src[2:, :] = get_3rd_point(src[0, :], src[1, :])
+    dst[2:, :] = get_3rd_point(dst[0, :], dst[1, :])
+
+    if inv:
+        trans = cv2.getAffineTransform(np.float32(dst), np.float32(src))
+    else:
+        trans = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return trans
+
+
+def affine_transform(pt, t):
+    new_pt = np.array([pt[0], pt[1], 1.0]).T
+    new_pt = np.dot(t, new_pt)
+    return new_pt[:2]
+
+
+def get_3rd_point(a, b):
+    direct = a - b
+    return b + np.array([-direct[1], direct[0]], dtype=np.float32)
+
+
+def get_dir(src_point, rot_rad):
+    sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+
+    src_result = [0, 0]
+    src_result[0] = src_point[0] * cs - src_point[1] * sn
+    src_result[1] = src_point[0] * sn + src_point[1] * cs
+
+    return src_result
+
+
+def crop(img, center, scale, output_size, rot=0):
+    trans = get_affine_transform(center, scale, rot, output_size)
+
+    dst_img = cv2.warpAffine(
+        img, trans, (int(output_size[1]), int(output_size[0])), flags=cv2.INTER_LINEAR
+    )
+
+    return dst_img

3rdparty/densepose/__init__.py

@@ -0,0 +1,20 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from .data.datasets import builtin  # just to register data
+from .converters import builtin as builtin_converters  # register converters
+from .config import (
+    add_densepose_config,
+    add_densepose_head_config,
+    add_hrnet_config,
+    add_dataset_category_config,
+    add_bootstrap_config,
+    load_bootstrap_config,
+)
+from .structures import DensePoseDataRelative, DensePoseList, DensePoseTransformData
+from .evaluation import DensePoseCOCOEvaluator
+from .modeling.roi_heads import DensePoseROIHeads
+from .modeling.test_time_augmentation import (
+    DensePoseGeneralizedRCNNWithTTA,
+    DensePoseDatasetMapperTTA,
+)
+from .utils.transform import load_from_cfg
+from .modeling.hrfpn import build_hrfpn_backbone

3rdparty/densepose/config.py

@@ -0,0 +1,277 @@
+# -*- coding = utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+# pyre-ignore-all-errors
+
+from detectron2.config import CfgNode as CN
+
+
+def add_dataset_category_config(cfg: CN) -> None:
+    """
+    Add config for additional category-related dataset options
+     - category whitelisting
+     - category mapping
+    """
+    _C = cfg
+    _C.DATASETS.CATEGORY_MAPS = CN(new_allowed=True)
+    _C.DATASETS.WHITELISTED_CATEGORIES = CN(new_allowed=True)
+    # class to mesh mapping
+    _C.DATASETS.CLASS_TO_MESH_NAME_MAPPING = CN(new_allowed=True)
+
+
+def add_evaluation_config(cfg: CN) -> None:
+    _C = cfg
+    _C.DENSEPOSE_EVALUATION = CN()
+    # evaluator type, possible values:
+    #  - "iou": evaluator for models that produce iou data
+    #  - "cse": evaluator for models that produce cse data
+    _C.DENSEPOSE_EVALUATION.TYPE = "iou"
+    # storage for DensePose results, possible values:
+    #  - "none": no explicit storage, all the results are stored in the
+    #            dictionary with predictions, memory intensive;
+    #            historically the default storage type
+    #  - "ram": RAM storage, uses per-process RAM storage, which is
+    #           reduced to a single process storage on later stages,
+    #           less memory intensive
+    #  - "file": file storage, uses per-process file-based storage,
+    #            the least memory intensive, but may create bottlenecks
+    #            on file system accesses
+    _C.DENSEPOSE_EVALUATION.STORAGE = "none"
+    # minimum threshold for IOU values: the lower its values is,
+    # the more matches are produced (and the higher the AP score)
+    _C.DENSEPOSE_EVALUATION.MIN_IOU_THRESHOLD = 0.5
+    # Non-distributed inference is slower (at inference time) but can avoid RAM OOM
+    _C.DENSEPOSE_EVALUATION.DISTRIBUTED_INFERENCE = True
+    # evaluate mesh alignment based on vertex embeddings, only makes sense in CSE context
+    _C.DENSEPOSE_EVALUATION.EVALUATE_MESH_ALIGNMENT = False
+    # meshes to compute mesh alignment for
+    _C.DENSEPOSE_EVALUATION.MESH_ALIGNMENT_MESH_NAMES = []
+
+
+def add_bootstrap_config(cfg: CN) -> None:
+    """ """
+    _C = cfg
+    _C.BOOTSTRAP_DATASETS = []
+    _C.BOOTSTRAP_MODEL = CN()
+    _C.BOOTSTRAP_MODEL.WEIGHTS = ""
+    _C.BOOTSTRAP_MODEL.DEVICE = "cuda"
+
+
+def get_bootstrap_dataset_config() -> CN:
+    _C = CN()
+    _C.DATASET = ""
+    # ratio used to mix data loaders
+    _C.RATIO = 0.1
+    # image loader
+    _C.IMAGE_LOADER = CN(new_allowed=True)
+    _C.IMAGE_LOADER.TYPE = ""
+    _C.IMAGE_LOADER.BATCH_SIZE = 4
+    _C.IMAGE_LOADER.NUM_WORKERS = 4
+    _C.IMAGE_LOADER.CATEGORIES = []
+    _C.IMAGE_LOADER.MAX_COUNT_PER_CATEGORY = 1_000_000
+    _C.IMAGE_LOADER.CATEGORY_TO_CLASS_MAPPING = CN(new_allowed=True)
+    # inference
+    _C.INFERENCE = CN()
+    # batch size for model inputs
+    _C.INFERENCE.INPUT_BATCH_SIZE = 4
+    # batch size to group model outputs
+    _C.INFERENCE.OUTPUT_BATCH_SIZE = 2
+    # sampled data
+    _C.DATA_SAMPLER = CN(new_allowed=True)
+    _C.DATA_SAMPLER.TYPE = ""
+    _C.DATA_SAMPLER.USE_GROUND_TRUTH_CATEGORIES = False
+    # filter
+    _C.FILTER = CN(new_allowed=True)
+    _C.FILTER.TYPE = ""
+    return _C
+
+
+def load_bootstrap_config(cfg: CN) -> None:
+    """
+    Bootstrap datasets are given as a list of `dict` that are not automatically
+    converted into CfgNode. This method processes all bootstrap dataset entries
+    and ensures that they are in CfgNode format and comply with the specification
+    """
+    if not cfg.BOOTSTRAP_DATASETS:
+        return
+
+    bootstrap_datasets_cfgnodes = []
+    for dataset_cfg in cfg.BOOTSTRAP_DATASETS:
+        _C = get_bootstrap_dataset_config().clone()
+        _C.merge_from_other_cfg(CN(dataset_cfg))
+        bootstrap_datasets_cfgnodes.append(_C)
+    cfg.BOOTSTRAP_DATASETS = bootstrap_datasets_cfgnodes
+
+
+def add_densepose_head_cse_config(cfg: CN) -> None:
+    """
+    Add configuration options for Continuous Surface Embeddings (CSE)
+    """
+    _C = cfg
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE = CN()
+    # Dimensionality D of the embedding space
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBED_SIZE = 16
+    # Embedder specifications for various mesh IDs
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBEDDERS = CN(new_allowed=True)
+    # normalization coefficient for embedding distances
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBEDDING_DIST_GAUSS_SIGMA = 0.01
+    # normalization coefficient for geodesic distances
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.GEODESIC_DIST_GAUSS_SIGMA = 0.01
+    # embedding loss weight
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBED_LOSS_WEIGHT = 0.6
+    # embedding loss name, currently the following options are supported:
+    # - EmbeddingLoss: cross-entropy on vertex labels
+    # - SoftEmbeddingLoss: cross-entropy on vertex label combined with
+    #    Gaussian penalty on distance between vertices
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBED_LOSS_NAME = "EmbeddingLoss"
+    # optimizer hyperparameters
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.FEATURES_LR_FACTOR = 1.0
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBEDDING_LR_FACTOR = 1.0
+    # Shape to shape cycle consistency loss parameters:
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.SHAPE_TO_SHAPE_CYCLE_LOSS = CN({"ENABLED": False})
+    # shape to shape cycle consistency loss weight
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.SHAPE_TO_SHAPE_CYCLE_LOSS.WEIGHT = 0.025
+    # norm type used for loss computation
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.SHAPE_TO_SHAPE_CYCLE_LOSS.NORM_P = 2
+    # normalization term for embedding similarity matrices
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.SHAPE_TO_SHAPE_CYCLE_LOSS.TEMPERATURE = 0.05
+    # maximum number of vertices to include into shape to shape cycle loss
+    # if negative or zero, all vertices are considered
+    # if positive, random subset of vertices of given size is considered
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.SHAPE_TO_SHAPE_CYCLE_LOSS.MAX_NUM_VERTICES = 4936
+    # Pixel to shape cycle consistency loss parameters:
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS = CN({"ENABLED": False})
+    # pixel to shape cycle consistency loss weight
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.WEIGHT = 0.0001
+    # norm type used for loss computation
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.NORM_P = 2
+    # map images to all meshes and back (if false, use only gt meshes from the batch)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.USE_ALL_MESHES_NOT_GT_ONLY = False
+    # Randomly select at most this number of pixels from every instance
+    # if negative or zero, all vertices are considered
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.NUM_PIXELS_TO_SAMPLE = 100
+    # normalization factor for pixel to pixel distances (higher value = smoother distribution)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.PIXEL_SIGMA = 5.0
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.TEMPERATURE_PIXEL_TO_VERTEX = 0.05
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CSE.PIX_TO_SHAPE_CYCLE_LOSS.TEMPERATURE_VERTEX_TO_PIXEL = 0.05
+
+
+def add_densepose_head_config(cfg: CN) -> None:
+    """
+    Add config for densepose head.
+    """
+    _C = cfg
+
+    _C.MODEL.DENSEPOSE_ON = True
+
+    _C.MODEL.ROI_DENSEPOSE_HEAD = CN()
+    _C.MODEL.ROI_DENSEPOSE_HEAD.NAME = ""
+    _C.MODEL.ROI_DENSEPOSE_HEAD.NUM_STACKED_CONVS = 8
+    # Number of parts used for point labels
+    _C.MODEL.ROI_DENSEPOSE_HEAD.NUM_PATCHES = 24
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECONV_KERNEL = 4
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CONV_HEAD_DIM = 512
+    _C.MODEL.ROI_DENSEPOSE_HEAD.CONV_HEAD_KERNEL = 3
+    _C.MODEL.ROI_DENSEPOSE_HEAD.UP_SCALE = 2
+    _C.MODEL.ROI_DENSEPOSE_HEAD.HEATMAP_SIZE = 112
+    _C.MODEL.ROI_DENSEPOSE_HEAD.POOLER_TYPE = "ROIAlignV2"
+    _C.MODEL.ROI_DENSEPOSE_HEAD.POOLER_RESOLUTION = 28
+    _C.MODEL.ROI_DENSEPOSE_HEAD.POOLER_SAMPLING_RATIO = 2
+    _C.MODEL.ROI_DENSEPOSE_HEAD.NUM_COARSE_SEGM_CHANNELS = 2  # 15 or 2
+    # Overlap threshold for an RoI to be considered foreground (if >= FG_IOU_THRESHOLD)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.FG_IOU_THRESHOLD = 0.7
+    # Loss weights for annotation masks.(14 Parts)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.INDEX_WEIGHTS = 5.0
+    # Loss weights for surface parts. (24 Parts)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.PART_WEIGHTS = 1.0
+    # Loss weights for UV regression.
+    _C.MODEL.ROI_DENSEPOSE_HEAD.POINT_REGRESSION_WEIGHTS = 0.01
+    # Coarse segmentation is trained using instance segmentation task data
+    _C.MODEL.ROI_DENSEPOSE_HEAD.COARSE_SEGM_TRAINED_BY_MASKS = False
+    # For Decoder
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECODER_ON = True
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECODER_NUM_CLASSES = 256
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECODER_CONV_DIMS = 256
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECODER_NORM = ""
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DECODER_COMMON_STRIDE = 4
+    # For DeepLab head
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DEEPLAB = CN()
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DEEPLAB.NORM = "GN"
+    _C.MODEL.ROI_DENSEPOSE_HEAD.DEEPLAB.NONLOCAL_ON = 0
+    # Predictor class name, must be registered in DENSEPOSE_PREDICTOR_REGISTRY
+    # Some registered predictors:
+    #   "DensePoseChartPredictor": predicts segmentation and UV coordinates for predefined charts
+    #   "DensePoseChartWithConfidencePredictor": predicts segmentation, UV coordinates
+    #       and associated confidences for predefined charts (default)
+    #   "DensePoseEmbeddingWithConfidencePredictor": predicts segmentation, embeddings
+    #       and associated confidences for CSE
+    _C.MODEL.ROI_DENSEPOSE_HEAD.PREDICTOR_NAME = "DensePoseChartWithConfidencePredictor"
+    # Loss class name, must be registered in DENSEPOSE_LOSS_REGISTRY
+    # Some registered losses:
+    #   "DensePoseChartLoss": loss for chart-based models that estimate
+    #      segmentation and UV coordinates
+    #   "DensePoseChartWithConfidenceLoss": loss for chart-based models that estimate
+    #      segmentation, UV coordinates and the corresponding confidences (default)
+    _C.MODEL.ROI_DENSEPOSE_HEAD.LOSS_NAME = "DensePoseChartWithConfidenceLoss"
+    # Confidences
+    # Enable learning UV confidences (variances) along with the actual values
+    _C.MODEL.ROI_DENSEPOSE_HEAD.UV_CONFIDENCE = CN({"ENABLED": False})
+    # UV confidence lower bound
+    _C.MODEL.ROI_DENSEPOSE_HEAD.UV_CONFIDENCE.EPSILON = 0.01
+    # Enable learning segmentation confidences (variances) along with the actual values
+    _C.MODEL.ROI_DENSEPOSE_HEAD.SEGM_CONFIDENCE = CN({"ENABLED": False})
+    # Segmentation confidence lower bound
+    _C.MODEL.ROI_DENSEPOSE_HEAD.SEGM_CONFIDENCE.EPSILON = 0.01
+    # Statistical model type for confidence learning, possible values:
+    # - "iid_iso": statistically independent identically distributed residuals
+    #    with isotropic covariance
+    # - "indep_aniso": statistically independent residuals with anisotropic
+    #    covariances
+    _C.MODEL.ROI_DENSEPOSE_HEAD.UV_CONFIDENCE.TYPE = "iid_iso"
+    # List of angles for rotation in data augmentation during training
+    _C.INPUT.ROTATION_ANGLES = [0]
+    _C.TEST.AUG.ROTATION_ANGLES = ()  # Rotation TTA
+
+    add_densepose_head_cse_config(cfg)
+
+
+def add_hrnet_config(cfg: CN) -> None:
+    """
+    Add config for HRNet backbone.
+    """
+    _C = cfg
+
+    # For HigherHRNet w32
+    _C.MODEL.HRNET = CN()
+    _C.MODEL.HRNET.STEM_INPLANES = 64
+    _C.MODEL.HRNET.STAGE2 = CN()
+    _C.MODEL.HRNET.STAGE2.NUM_MODULES = 1
+    _C.MODEL.HRNET.STAGE2.NUM_BRANCHES = 2
+    _C.MODEL.HRNET.STAGE2.BLOCK = "BASIC"
+    _C.MODEL.HRNET.STAGE2.NUM_BLOCKS = [4, 4]
+    _C.MODEL.HRNET.STAGE2.NUM_CHANNELS = [32, 64]
+    _C.MODEL.HRNET.STAGE2.FUSE_METHOD = "SUM"
+    _C.MODEL.HRNET.STAGE3 = CN()
+    _C.MODEL.HRNET.STAGE3.NUM_MODULES = 4
+    _C.MODEL.HRNET.STAGE3.NUM_BRANCHES = 3
+    _C.MODEL.HRNET.STAGE3.BLOCK = "BASIC"
+    _C.MODEL.HRNET.STAGE3.NUM_BLOCKS = [4, 4, 4]
+    _C.MODEL.HRNET.STAGE3.NUM_CHANNELS = [32, 64, 128]
+    _C.MODEL.HRNET.STAGE3.FUSE_METHOD = "SUM"
+    _C.MODEL.HRNET.STAGE4 = CN()
+    _C.MODEL.HRNET.STAGE4.NUM_MODULES = 3
+    _C.MODEL.HRNET.STAGE4.NUM_BRANCHES = 4
+    _C.MODEL.HRNET.STAGE4.BLOCK = "BASIC"
+    _C.MODEL.HRNET.STAGE4.NUM_BLOCKS = [4, 4, 4, 4]
+    _C.MODEL.HRNET.STAGE4.NUM_CHANNELS = [32, 64, 128, 256]
+    _C.MODEL.HRNET.STAGE4.FUSE_METHOD = "SUM"
+
+    _C.MODEL.HRNET.HRFPN = CN()
+    _C.MODEL.HRNET.HRFPN.OUT_CHANNELS = 256
+
+
+def add_densepose_config(cfg: CN) -> None:
+    add_densepose_head_config(cfg)
+    add_hrnet_config(cfg)
+    add_bootstrap_config(cfg)
+    add_dataset_category_config(cfg)
+    add_evaluation_config(cfg)

3rdparty/densepose/converters/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .hflip import HFlipConverter
+from .to_mask import ToMaskConverter
+from .to_chart_result import ToChartResultConverter, ToChartResultConverterWithConfidences
+from .segm_to_mask import (
+    predictor_output_with_fine_and_coarse_segm_to_mask,
+    predictor_output_with_coarse_segm_to_mask,
+    resample_fine_and_coarse_segm_to_bbox,
+)
+from .chart_output_to_chart_result import (
+    densepose_chart_predictor_output_to_result,
+    densepose_chart_predictor_output_to_result_with_confidences,
+)
+from .chart_output_hflip import densepose_chart_predictor_output_hflip

3rdparty/densepose/converters/base.py

@@ -0,0 +1,93 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any, Tuple, Type
+import torch
+
+
+class BaseConverter:
+    """
+    Converter base class to be reused by various converters.
+    Converter allows one to convert data from various source types to a particular
+    destination type. Each source type needs to register its converter. The
+    registration for each source type is valid for all descendants of that type.
+    """
+
+    @classmethod
+    def register(cls, from_type: Type, converter: Any = None):
+        """
+        Registers a converter for the specified type.
+        Can be used as a decorator (if converter is None), or called as a method.
+
+        Args:
+            from_type (type): type to register the converter for;
+                all instances of this type will use the same converter
+            converter (callable): converter to be registered for the given
+                type; if None, this method is assumed to be a decorator for the converter
+        """
+
+        if converter is not None:
+            cls._do_register(from_type, converter)
+
+        def wrapper(converter: Any) -> Any:
+            cls._do_register(from_type, converter)
+            return converter
+
+        return wrapper
+
+    @classmethod
+    def _do_register(cls, from_type: Type, converter: Any):
+        cls.registry[from_type] = converter  # pyre-ignore[16]
+
+    @classmethod
+    def _lookup_converter(cls, from_type: Type) -> Any:
+        """
+        Perform recursive lookup for the given type
+        to find registered converter. If a converter was found for some base
+        class, it gets registered for this class to save on further lookups.
+
+        Args:
+            from_type: type for which to find a converter
+        Return:
+            callable or None - registered converter or None
+                if no suitable entry was found in the registry
+        """
+        if from_type in cls.registry:  # pyre-ignore[16]
+            return cls.registry[from_type]
+        for base in from_type.__bases__:
+            converter = cls._lookup_converter(base)
+            if converter is not None:
+                cls._do_register(from_type, converter)
+                return converter
+        return None
+
+    @classmethod
+    def convert(cls, instance: Any, *args, **kwargs):
+        """
+        Convert an instance to the destination type using some registered
+        converter. Does recursive lookup for base classes, so there's no need
+        for explicit registration for derived classes.
+
+        Args:
+            instance: source instance to convert to the destination type
+        Return:
+            An instance of the destination type obtained from the source instance
+            Raises KeyError, if no suitable converter found
+        """
+        instance_type = type(instance)
+        converter = cls._lookup_converter(instance_type)
+        if converter is None:
+            if cls.dst_type is None:  # pyre-ignore[16]
+                output_type_str = "itself"
+            else:
+                output_type_str = cls.dst_type
+            raise KeyError(f"Could not find converter from {instance_type} to {output_type_str}")
+        return converter(instance, *args, **kwargs)
+
+
+IntTupleBox = Tuple[int, int, int, int]
+
+
+def make_int_box(box: torch.Tensor) -> IntTupleBox:
+    int_box = [0, 0, 0, 0]
+    int_box[0], int_box[1], int_box[2], int_box[3] = tuple(box.long().tolist())
+    return int_box[0], int_box[1], int_box[2], int_box[3]

3rdparty/densepose/converters/builtin.py

@@ -0,0 +1,31 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from ..structures import DensePoseChartPredictorOutput, DensePoseEmbeddingPredictorOutput
+from . import (
+    HFlipConverter,
+    ToChartResultConverter,
+    ToChartResultConverterWithConfidences,
+    ToMaskConverter,
+    densepose_chart_predictor_output_hflip,
+    densepose_chart_predictor_output_to_result,
+    densepose_chart_predictor_output_to_result_with_confidences,
+    predictor_output_with_coarse_segm_to_mask,
+    predictor_output_with_fine_and_coarse_segm_to_mask,
+)
+
+ToMaskConverter.register(
+    DensePoseChartPredictorOutput, predictor_output_with_fine_and_coarse_segm_to_mask
+)
+ToMaskConverter.register(
+    DensePoseEmbeddingPredictorOutput, predictor_output_with_coarse_segm_to_mask
+)
+
+ToChartResultConverter.register(
+    DensePoseChartPredictorOutput, densepose_chart_predictor_output_to_result
+)
+
+ToChartResultConverterWithConfidences.register(
+    DensePoseChartPredictorOutput, densepose_chart_predictor_output_to_result_with_confidences
+)
+
+HFlipConverter.register(DensePoseChartPredictorOutput, densepose_chart_predictor_output_hflip)

3rdparty/densepose/converters/chart_output_hflip.py

@@ -0,0 +1,71 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from dataclasses import fields
+import torch
+
+from densepose.structures import DensePoseChartPredictorOutput, DensePoseTransformData
+
+
+def densepose_chart_predictor_output_hflip(
+    densepose_predictor_output: DensePoseChartPredictorOutput,
+    transform_data: DensePoseTransformData,
+) -> DensePoseChartPredictorOutput:
+    """
+    Change  to take into account a Horizontal flip.
+    """
+    if len(densepose_predictor_output) > 0:
+
+        PredictorOutput = type(densepose_predictor_output)
+        output_dict = {}
+
+        for field in fields(densepose_predictor_output):
+            field_value = getattr(densepose_predictor_output, field.name)
+            # flip tensors
+            if isinstance(field_value, torch.Tensor):
+                setattr(densepose_predictor_output, field.name, torch.flip(field_value, [3]))
+
+        densepose_predictor_output = _flip_iuv_semantics_tensor(
+            densepose_predictor_output, transform_data
+        )
+        densepose_predictor_output = _flip_segm_semantics_tensor(
+            densepose_predictor_output, transform_data
+        )
+
+        for field in fields(densepose_predictor_output):
+            output_dict[field.name] = getattr(densepose_predictor_output, field.name)
+
+        return PredictorOutput(**output_dict)
+    else:
+        return densepose_predictor_output
+
+
+def _flip_iuv_semantics_tensor(
+    densepose_predictor_output: DensePoseChartPredictorOutput,
+    dp_transform_data: DensePoseTransformData,
+) -> DensePoseChartPredictorOutput:
+    point_label_symmetries = dp_transform_data.point_label_symmetries
+    uv_symmetries = dp_transform_data.uv_symmetries
+
+    N, C, H, W = densepose_predictor_output.u.shape
+    u_loc = (densepose_predictor_output.u[:, 1:, :, :].clamp(0, 1) * 255).long()
+    v_loc = (densepose_predictor_output.v[:, 1:, :, :].clamp(0, 1) * 255).long()
+    Iindex = torch.arange(C - 1, device=densepose_predictor_output.u.device)[
+        None, :, None, None
+    ].expand(N, C - 1, H, W)
+    densepose_predictor_output.u[:, 1:, :, :] = uv_symmetries["U_transforms"][Iindex, v_loc, u_loc]
+    densepose_predictor_output.v[:, 1:, :, :] = uv_symmetries["V_transforms"][Iindex, v_loc, u_loc]
+
+    for el in ["fine_segm", "u", "v"]:
+        densepose_predictor_output.__dict__[el] = densepose_predictor_output.__dict__[el][
+            :, point_label_symmetries, :, :
+        ]
+    return densepose_predictor_output
+
+
+def _flip_segm_semantics_tensor(
+    densepose_predictor_output: DensePoseChartPredictorOutput, dp_transform_data
+):
+    if densepose_predictor_output.coarse_segm.shape[1] > 2:
+        densepose_predictor_output.coarse_segm = densepose_predictor_output.coarse_segm[
+            :, dp_transform_data.mask_label_symmetries, :, :
+        ]
+    return densepose_predictor_output

3rdparty/densepose/converters/chart_output_to_chart_result.py

@@ -0,0 +1,188 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Dict
+import torch
+from torch.nn import functional as F
+
+from detectron2.structures.boxes import Boxes, BoxMode
+
+from ..structures import (
+    DensePoseChartPredictorOutput,
+    DensePoseChartResult,
+    DensePoseChartResultWithConfidences,
+)
+from . import resample_fine_and_coarse_segm_to_bbox
+from .base import IntTupleBox, make_int_box
+
+
+def resample_uv_tensors_to_bbox(
+    u: torch.Tensor,
+    v: torch.Tensor,
+    labels: torch.Tensor,
+    box_xywh_abs: IntTupleBox,
+) -> torch.Tensor:
+    """
+    Resamples U and V coordinate estimates for the given bounding box
+
+    Args:
+        u (tensor [1, C, H, W] of float): U coordinates
+        v (tensor [1, C, H, W] of float): V coordinates
+        labels (tensor [H, W] of long): labels obtained by resampling segmentation
+            outputs for the given bounding box
+        box_xywh_abs (tuple of 4 int): bounding box that corresponds to predictor outputs
+    Return:
+       Resampled U and V coordinates - a tensor [2, H, W] of float
+    """
+    x, y, w, h = box_xywh_abs
+    w = max(int(w), 1)
+    h = max(int(h), 1)
+    u_bbox = F.interpolate(u, (h, w), mode="bilinear", align_corners=False)
+    v_bbox = F.interpolate(v, (h, w), mode="bilinear", align_corners=False)
+    uv = torch.zeros([2, h, w], dtype=torch.float32, device=u.device)
+    for part_id in range(1, u_bbox.size(1)):
+        uv[0][labels == part_id] = u_bbox[0, part_id][labels == part_id]
+        uv[1][labels == part_id] = v_bbox[0, part_id][labels == part_id]
+    return uv
+
+
+def resample_uv_to_bbox(
+    predictor_output: DensePoseChartPredictorOutput,
+    labels: torch.Tensor,
+    box_xywh_abs: IntTupleBox,
+) -> torch.Tensor:
+    """
+    Resamples U and V coordinate estimates for the given bounding box
+
+    Args:
+        predictor_output (DensePoseChartPredictorOutput): DensePose predictor
+            output to be resampled
+        labels (tensor [H, W] of long): labels obtained by resampling segmentation
+            outputs for the given bounding box
+        box_xywh_abs (tuple of 4 int): bounding box that corresponds to predictor outputs
+    Return:
+       Resampled U and V coordinates - a tensor [2, H, W] of float
+    """
+    return resample_uv_tensors_to_bbox(
+        predictor_output.u,
+        predictor_output.v,
+        labels,
+        box_xywh_abs,
+    )
+
+
+def densepose_chart_predictor_output_to_result(
+    predictor_output: DensePoseChartPredictorOutput, boxes: Boxes
+) -> DensePoseChartResult:
+    """
+    Convert densepose chart predictor outputs to results
+
+    Args:
+        predictor_output (DensePoseChartPredictorOutput): DensePose predictor
+            output to be converted to results, must contain only 1 output
+        boxes (Boxes): bounding box that corresponds to the predictor output,
+            must contain only 1 bounding box
+    Return:
+       DensePose chart-based result (DensePoseChartResult)
+    """
+    assert len(predictor_output) == 1 and len(boxes) == 1, (
+        f"Predictor output to result conversion can operate only single outputs"
+        f", got {len(predictor_output)} predictor outputs and {len(boxes)} boxes"
+    )
+
+    boxes_xyxy_abs = boxes.tensor.clone()
+    boxes_xywh_abs = BoxMode.convert(boxes_xyxy_abs, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+    box_xywh = make_int_box(boxes_xywh_abs[0])
+
+    labels = resample_fine_and_coarse_segm_to_bbox(predictor_output, box_xywh).squeeze(0)
+    uv = resample_uv_to_bbox(predictor_output, labels, box_xywh)
+    return DensePoseChartResult(labels=labels, uv=uv)
+
+
+def resample_confidences_to_bbox(
+    predictor_output: DensePoseChartPredictorOutput,
+    labels: torch.Tensor,
+    box_xywh_abs: IntTupleBox,
+) -> Dict[str, torch.Tensor]:
+    """
+    Resamples confidences for the given bounding box
+
+    Args:
+        predictor_output (DensePoseChartPredictorOutput): DensePose predictor
+            output to be resampled
+        labels (tensor [H, W] of long): labels obtained by resampling segmentation
+            outputs for the given bounding box
+        box_xywh_abs (tuple of 4 int): bounding box that corresponds to predictor outputs
+    Return:
+       Resampled confidences - a dict of [H, W] tensors of float
+    """
+
+    x, y, w, h = box_xywh_abs
+    w = max(int(w), 1)
+    h = max(int(h), 1)
+
+    confidence_names = [
+        "sigma_1",
+        "sigma_2",
+        "kappa_u",
+        "kappa_v",
+        "fine_segm_confidence",
+        "coarse_segm_confidence",
+    ]
+    confidence_results = {key: None for key in confidence_names}
+    confidence_names = [
+        key for key in confidence_names if getattr(predictor_output, key) is not None
+    ]
+    confidence_base = torch.zeros([h, w], dtype=torch.float32, device=predictor_output.u.device)
+
+    # assign data from channels that correspond to the labels
+    for key in confidence_names:
+        resampled_confidence = F.interpolate(
+            getattr(predictor_output, key),
+            (h, w),
+            mode="bilinear",
+            align_corners=False,
+        )
+        result = confidence_base.clone()
+        for part_id in range(1, predictor_output.u.size(1)):
+            if resampled_confidence.size(1) != predictor_output.u.size(1):
+                # confidence is not part-based, don't try to fill it part by part
+                continue
+            result[labels == part_id] = resampled_confidence[0, part_id][labels == part_id]
+
+        if resampled_confidence.size(1) != predictor_output.u.size(1):
+            # confidence is not part-based, fill the data with the first channel
+            # (targeted for segmentation confidences that have only 1 channel)
+            result = resampled_confidence[0, 0]
+
+        confidence_results[key] = result
+
+    return confidence_results  # pyre-ignore[7]
+
+
+def densepose_chart_predictor_output_to_result_with_confidences(
+    predictor_output: DensePoseChartPredictorOutput, boxes: Boxes
+) -> DensePoseChartResultWithConfidences:
+    """
+    Convert densepose chart predictor outputs to results
+
+    Args:
+        predictor_output (DensePoseChartPredictorOutput): DensePose predictor
+            output with confidences to be converted to results, must contain only 1 output
+        boxes (Boxes): bounding box that corresponds to the predictor output,
+            must contain only 1 bounding box
+    Return:
+       DensePose chart-based result with confidences (DensePoseChartResultWithConfidences)
+    """
+    assert len(predictor_output) == 1 and len(boxes) == 1, (
+        f"Predictor output to result conversion can operate only single outputs"
+        f", got {len(predictor_output)} predictor outputs and {len(boxes)} boxes"
+    )
+
+    boxes_xyxy_abs = boxes.tensor.clone()
+    boxes_xywh_abs = BoxMode.convert(boxes_xyxy_abs, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+    box_xywh = make_int_box(boxes_xywh_abs[0])
+
+    labels = resample_fine_and_coarse_segm_to_bbox(predictor_output, box_xywh).squeeze(0)
+    uv = resample_uv_to_bbox(predictor_output, labels, box_xywh)
+    confidences = resample_confidences_to_bbox(predictor_output, labels, box_xywh)
+    return DensePoseChartResultWithConfidences(labels=labels, uv=uv, **confidences)

3rdparty/densepose/converters/hflip.py

@@ -0,0 +1,34 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any
+
+from .base import BaseConverter
+
+
+class HFlipConverter(BaseConverter):
+    """
+    Converts various DensePose predictor outputs to DensePose results.
+    Each DensePose predictor output type has to register its convertion strategy.
+    """
+
+    registry = {}
+    dst_type = None
+
+    @classmethod
+    # pyre-fixme[14]: `convert` overrides method defined in `BaseConverter`
+    #  inconsistently.
+    def convert(cls, predictor_outputs: Any, transform_data: Any, *args, **kwargs):
+        """
+        Performs an horizontal flip on DensePose predictor outputs.
+        Does recursive lookup for base classes, so there's no need
+        for explicit registration for derived classes.
+
+        Args:
+            predictor_outputs: DensePose predictor output to be converted to BitMasks
+            transform_data: Anything useful for the flip
+        Return:
+            An instance of the same type as predictor_outputs
+        """
+        return super(HFlipConverter, cls).convert(
+            predictor_outputs, transform_data, *args, **kwargs
+        )

3rdparty/densepose/converters/segm_to_mask.py

@@ -0,0 +1,150 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any
+import torch
+from torch.nn import functional as F
+
+from detectron2.structures import BitMasks, Boxes, BoxMode
+
+from .base import IntTupleBox, make_int_box
+from .to_mask import ImageSizeType
+
+
+def resample_coarse_segm_tensor_to_bbox(coarse_segm: torch.Tensor, box_xywh_abs: IntTupleBox):
+    """
+    Resample coarse segmentation tensor to the given
+    bounding box and derive labels for each pixel of the bounding box
+
+    Args:
+        coarse_segm: float tensor of shape [1, K, Hout, Wout]
+        box_xywh_abs (tuple of 4 int): bounding box given by its upper-left
+            corner coordinates, width (W) and height (H)
+    Return:
+        Labels for each pixel of the bounding box, a long tensor of size [1, H, W]
+    """
+    x, y, w, h = box_xywh_abs
+    w = max(int(w), 1)
+    h = max(int(h), 1)
+    labels = F.interpolate(coarse_segm, (h, w), mode="bilinear", align_corners=False).argmax(dim=1)
+    return labels
+
+
+def resample_fine_and_coarse_segm_tensors_to_bbox(
+    fine_segm: torch.Tensor, coarse_segm: torch.Tensor, box_xywh_abs: IntTupleBox
+):
+    """
+    Resample fine and coarse segmentation tensors to the given
+    bounding box and derive labels for each pixel of the bounding box
+
+    Args:
+        fine_segm: float tensor of shape [1, C, Hout, Wout]
+        coarse_segm: float tensor of shape [1, K, Hout, Wout]
+        box_xywh_abs (tuple of 4 int): bounding box given by its upper-left
+            corner coordinates, width (W) and height (H)
+    Return:
+        Labels for each pixel of the bounding box, a long tensor of size [1, H, W]
+    """
+    x, y, w, h = box_xywh_abs
+    w = max(int(w), 1)
+    h = max(int(h), 1)
+    # coarse segmentation
+    coarse_segm_bbox = F.interpolate(
+        coarse_segm,
+        (h, w),
+        mode="bilinear",
+        align_corners=False,
+    ).argmax(dim=1)
+    # combined coarse and fine segmentation
+    labels = (
+        F.interpolate(fine_segm, (h, w), mode="bilinear", align_corners=False).argmax(dim=1)
+        * (coarse_segm_bbox > 0).long()
+    )
+    return labels
+
+
+def resample_fine_and_coarse_segm_to_bbox(predictor_output: Any, box_xywh_abs: IntTupleBox):
+    """
+    Resample fine and coarse segmentation outputs from a predictor to the given
+    bounding box and derive labels for each pixel of the bounding box
+
+    Args:
+        predictor_output: DensePose predictor output that contains segmentation
+            results to be resampled
+        box_xywh_abs (tuple of 4 int): bounding box given by its upper-left
+            corner coordinates, width (W) and height (H)
+    Return:
+        Labels for each pixel of the bounding box, a long tensor of size [1, H, W]
+    """
+    return resample_fine_and_coarse_segm_tensors_to_bbox(
+        predictor_output.fine_segm,
+        predictor_output.coarse_segm,
+        box_xywh_abs,
+    )
+
+
+def predictor_output_with_coarse_segm_to_mask(
+    predictor_output: Any, boxes: Boxes, image_size_hw: ImageSizeType
+) -> BitMasks:
+    """
+    Convert predictor output with coarse and fine segmentation to a mask.
+    Assumes that predictor output has the following attributes:
+     - coarse_segm (tensor of size [N, D, H, W]): coarse segmentation
+         unnormalized scores for N instances; D is the number of coarse
+         segmentation labels, H and W is the resolution of the estimate
+
+    Args:
+        predictor_output: DensePose predictor output to be converted to mask
+        boxes (Boxes): bounding boxes that correspond to the DensePose
+            predictor outputs
+        image_size_hw (tuple [int, int]): image height Himg and width Wimg
+    Return:
+        BitMasks that contain a bool tensor of size [N, Himg, Wimg] with
+        a mask of the size of the image for each instance
+    """
+    H, W = image_size_hw
+    boxes_xyxy_abs = boxes.tensor.clone()
+    boxes_xywh_abs = BoxMode.convert(boxes_xyxy_abs, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+    N = len(boxes_xywh_abs)
+    masks = torch.zeros((N, H, W), dtype=torch.bool, device=boxes.tensor.device)
+    for i in range(len(boxes_xywh_abs)):
+        box_xywh = make_int_box(boxes_xywh_abs[i])
+        box_mask = resample_coarse_segm_tensor_to_bbox(predictor_output[i].coarse_segm, box_xywh)
+        x, y, w, h = box_xywh
+        masks[i, y : y + h, x : x + w] = box_mask
+
+    return BitMasks(masks)
+
+
+def predictor_output_with_fine_and_coarse_segm_to_mask(
+    predictor_output: Any, boxes: Boxes, image_size_hw: ImageSizeType
+) -> BitMasks:
+    """
+    Convert predictor output with coarse and fine segmentation to a mask.
+    Assumes that predictor output has the following attributes:
+     - coarse_segm (tensor of size [N, D, H, W]): coarse segmentation
+         unnormalized scores for N instances; D is the number of coarse
+         segmentation labels, H and W is the resolution of the estimate
+     - fine_segm (tensor of size [N, C, H, W]): fine segmentation
+         unnormalized scores for N instances; C is the number of fine
+         segmentation labels, H and W is the resolution of the estimate
+
+    Args:
+        predictor_output: DensePose predictor output to be converted to mask
+        boxes (Boxes): bounding boxes that correspond to the DensePose
+            predictor outputs
+        image_size_hw (tuple [int, int]): image height Himg and width Wimg
+    Return:
+        BitMasks that contain a bool tensor of size [N, Himg, Wimg] with
+        a mask of the size of the image for each instance
+    """
+    H, W = image_size_hw
+    boxes_xyxy_abs = boxes.tensor.clone()
+    boxes_xywh_abs = BoxMode.convert(boxes_xyxy_abs, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+    N = len(boxes_xywh_abs)
+    masks = torch.zeros((N, H, W), dtype=torch.bool, device=boxes.tensor.device)
+    for i in range(len(boxes_xywh_abs)):
+        box_xywh = make_int_box(boxes_xywh_abs[i])
+        labels_i = resample_fine_and_coarse_segm_to_bbox(predictor_output[i], box_xywh)
+        x, y, w, h = box_xywh
+        masks[i, y : y + h, x : x + w] = labels_i > 0
+    return BitMasks(masks)

3rdparty/densepose/converters/to_chart_result.py

@@ -0,0 +1,70 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any
+
+from detectron2.structures import Boxes
+
+from ..structures import DensePoseChartResult, DensePoseChartResultWithConfidences
+from .base import BaseConverter
+
+
+class ToChartResultConverter(BaseConverter):
+    """
+    Converts various DensePose predictor outputs to DensePose results.
+    Each DensePose predictor output type has to register its convertion strategy.
+    """
+
+    registry = {}
+    dst_type = DensePoseChartResult
+
+    @classmethod
+    # pyre-fixme[14]: `convert` overrides method defined in `BaseConverter`
+    #  inconsistently.
+    def convert(cls, predictor_outputs: Any, boxes: Boxes, *args, **kwargs) -> DensePoseChartResult:
+        """
+        Convert DensePose predictor outputs to DensePoseResult using some registered
+        converter. Does recursive lookup for base classes, so there's no need
+        for explicit registration for derived classes.
+
+        Args:
+            densepose_predictor_outputs: DensePose predictor output to be
+                converted to BitMasks
+            boxes (Boxes): bounding boxes that correspond to the DensePose
+                predictor outputs
+        Return:
+            An instance of DensePoseResult. If no suitable converter was found, raises KeyError
+        """
+        return super(ToChartResultConverter, cls).convert(predictor_outputs, boxes, *args, **kwargs)
+
+
+class ToChartResultConverterWithConfidences(BaseConverter):
+    """
+    Converts various DensePose predictor outputs to DensePose results.
+    Each DensePose predictor output type has to register its convertion strategy.
+    """
+
+    registry = {}
+    dst_type = DensePoseChartResultWithConfidences
+
+    @classmethod
+    # pyre-fixme[14]: `convert` overrides method defined in `BaseConverter`
+    #  inconsistently.
+    def convert(
+        cls, predictor_outputs: Any, boxes: Boxes, *args, **kwargs
+    ) -> DensePoseChartResultWithConfidences:
+        """
+        Convert DensePose predictor outputs to DensePoseResult with confidences
+        using some registered converter. Does recursive lookup for base classes,
+        so there's no need for explicit registration for derived classes.
+
+        Args:
+            densepose_predictor_outputs: DensePose predictor output with confidences
+                to be converted to BitMasks
+            boxes (Boxes): bounding boxes that correspond to the DensePose
+                predictor outputs
+        Return:
+            An instance of DensePoseResult. If no suitable converter was found, raises KeyError
+        """
+        return super(ToChartResultConverterWithConfidences, cls).convert(
+            predictor_outputs, boxes, *args, **kwargs
+        )

3rdparty/densepose/converters/to_mask.py

@@ -0,0 +1,49 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any, Tuple
+
+from detectron2.structures import BitMasks, Boxes
+
+from .base import BaseConverter
+
+ImageSizeType = Tuple[int, int]
+
+
+class ToMaskConverter(BaseConverter):
+    """
+    Converts various DensePose predictor outputs to masks
+    in bit mask format (see `BitMasks`). Each DensePose predictor output type
+    has to register its convertion strategy.
+    """
+
+    registry = {}
+    dst_type = BitMasks
+
+    @classmethod
+    # pyre-fixme[14]: `convert` overrides method defined in `BaseConverter`
+    #  inconsistently.
+    def convert(
+        cls,
+        densepose_predictor_outputs: Any,
+        boxes: Boxes,
+        image_size_hw: ImageSizeType,
+        *args,
+        **kwargs
+    ) -> BitMasks:
+        """
+        Convert DensePose predictor outputs to BitMasks using some registered
+        converter. Does recursive lookup for base classes, so there's no need
+        for explicit registration for derived classes.
+
+        Args:
+            densepose_predictor_outputs: DensePose predictor output to be
+                converted to BitMasks
+            boxes (Boxes): bounding boxes that correspond to the DensePose
+                predictor outputs
+            image_size_hw (tuple [int, int]): image height and width
+        Return:
+            An instance of `BitMasks`. If no suitable converter was found, raises KeyError
+        """
+        return super(ToMaskConverter, cls).convert(
+            densepose_predictor_outputs, boxes, image_size_hw, *args, **kwargs
+        )

3rdparty/densepose/data/__init__.py

@@ -0,0 +1,25 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .meshes import builtin
+from .build import (
+    build_detection_test_loader,
+    build_detection_train_loader,
+    build_combined_loader,
+    build_frame_selector,
+    build_inference_based_loaders,
+    has_inference_based_loaders,
+    BootstrapDatasetFactoryCatalog,
+)
+from .combined_loader import CombinedDataLoader
+from .dataset_mapper import DatasetMapper
+from .inference_based_loader import InferenceBasedLoader, ScoreBasedFilter
+from .image_list_dataset import ImageListDataset
+from .utils import is_relative_local_path, maybe_prepend_base_path
+
+# ensure the builtin datasets are registered
+from . import datasets
+
+# ensure the bootstrap datasets builders are registered
+from . import build
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

3rdparty/densepose/data/build.py

@@ -0,0 +1,736 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import itertools
+import logging
+import numpy as np
+from collections import UserDict, defaultdict
+from dataclasses import dataclass
+from typing import Any, Callable, Collection, Dict, Iterable, List, Optional, Sequence, Tuple
+import torch
+from torch.utils.data.dataset import Dataset
+
+from detectron2.config import CfgNode
+from detectron2.data.build import build_detection_test_loader as d2_build_detection_test_loader
+from detectron2.data.build import build_detection_train_loader as d2_build_detection_train_loader
+from detectron2.data.build import (
+    load_proposals_into_dataset,
+    print_instances_class_histogram,
+    trivial_batch_collator,
+    worker_init_reset_seed,
+)
+from detectron2.data.catalog import DatasetCatalog, Metadata, MetadataCatalog
+from detectron2.data.samplers import TrainingSampler
+from detectron2.utils.comm import get_world_size
+
+from densepose.config import get_bootstrap_dataset_config
+from densepose.modeling import build_densepose_embedder
+
+from .combined_loader import CombinedDataLoader, Loader
+from .dataset_mapper import DatasetMapper
+from .datasets.coco import DENSEPOSE_CSE_KEYS_WITHOUT_MASK, DENSEPOSE_IUV_KEYS_WITHOUT_MASK
+from .datasets.dataset_type import DatasetType
+from .inference_based_loader import InferenceBasedLoader, ScoreBasedFilter
+from .samplers import (
+    DensePoseConfidenceBasedSampler,
+    DensePoseCSEConfidenceBasedSampler,
+    DensePoseCSEUniformSampler,
+    DensePoseUniformSampler,
+    MaskFromDensePoseSampler,
+    PredictionToGroundTruthSampler,
+)
+from .transform import ImageResizeTransform
+from .utils import get_category_to_class_mapping, get_class_to_mesh_name_mapping
+from .video import (
+    FirstKFramesSelector,
+    FrameSelectionStrategy,
+    LastKFramesSelector,
+    RandomKFramesSelector,
+    VideoKeyframeDataset,
+    video_list_from_file,
+)
+
+__all__ = ["build_detection_train_loader", "build_detection_test_loader"]
+
+
+Instance = Dict[str, Any]
+InstancePredicate = Callable[[Instance], bool]
+
+
+def _compute_num_images_per_worker(cfg: CfgNode) -> int:
+    num_workers = get_world_size()
+    images_per_batch = cfg.SOLVER.IMS_PER_BATCH
+    assert (
+        images_per_batch % num_workers == 0
+    ), "SOLVER.IMS_PER_BATCH ({}) must be divisible by the number of workers ({}).".format(
+        images_per_batch, num_workers
+    )
+    assert (
+        images_per_batch >= num_workers
+    ), "SOLVER.IMS_PER_BATCH ({}) must be larger than the number of workers ({}).".format(
+        images_per_batch, num_workers
+    )
+    images_per_worker = images_per_batch // num_workers
+    return images_per_worker
+
+
+def _map_category_id_to_contiguous_id(dataset_name: str, dataset_dicts: Iterable[Instance]) -> None:
+    meta = MetadataCatalog.get(dataset_name)
+    for dataset_dict in dataset_dicts:
+        for ann in dataset_dict["annotations"]:
+            ann["category_id"] = meta.thing_dataset_id_to_contiguous_id[ann["category_id"]]
+
+
+@dataclass
+class _DatasetCategory:
+    """
+    Class representing category data in a dataset:
+     - id: category ID, as specified in the dataset annotations file
+     - name: category name, as specified in the dataset annotations file
+     - mapped_id: category ID after applying category maps (DATASETS.CATEGORY_MAPS config option)
+     - mapped_name: category name after applying category maps
+     - dataset_name: dataset in which the category is defined
+
+    For example, when training models in a class-agnostic manner, one could take LVIS 1.0
+    dataset and map the animal categories to the same category as human data from COCO:
+     id = 225
+     name = "cat"
+     mapped_id = 1
+     mapped_name = "person"
+     dataset_name = "lvis_v1_animals_dp_train"
+    """
+
+    id: int
+    name: str
+    mapped_id: int
+    mapped_name: str
+    dataset_name: str
+
+
+_MergedCategoriesT = Dict[int, List[_DatasetCategory]]
+
+
+def _add_category_id_to_contiguous_id_maps_to_metadata(
+    merged_categories: _MergedCategoriesT,
+) -> None:
+    merged_categories_per_dataset = {}
+    for contiguous_cat_id, cat_id in enumerate(sorted(merged_categories.keys())):
+        for cat in merged_categories[cat_id]:
+            if cat.dataset_name not in merged_categories_per_dataset:
+                merged_categories_per_dataset[cat.dataset_name] = defaultdict(list)
+            merged_categories_per_dataset[cat.dataset_name][cat_id].append(
+                (
+                    contiguous_cat_id,
+                    cat,
+                )
+            )
+
+    logger = logging.getLogger(__name__)
+    for dataset_name, merged_categories in merged_categories_per_dataset.items():
+        meta = MetadataCatalog.get(dataset_name)
+        if not hasattr(meta, "thing_classes"):
+            meta.thing_classes = []
+            meta.thing_dataset_id_to_contiguous_id = {}
+            meta.thing_dataset_id_to_merged_id = {}
+        else:
+            meta.thing_classes.clear()
+            meta.thing_dataset_id_to_contiguous_id.clear()
+            meta.thing_dataset_id_to_merged_id.clear()
+        logger.info(f"Dataset {dataset_name}: category ID to contiguous ID mapping:")
+        for _cat_id, categories in sorted(merged_categories.items()):
+            added_to_thing_classes = False
+            for contiguous_cat_id, cat in categories:
+                if not added_to_thing_classes:
+                    meta.thing_classes.append(cat.mapped_name)
+                    added_to_thing_classes = True
+                meta.thing_dataset_id_to_contiguous_id[cat.id] = contiguous_cat_id
+                meta.thing_dataset_id_to_merged_id[cat.id] = cat.mapped_id
+                logger.info(f"{cat.id} ({cat.name}) -> {contiguous_cat_id}")
+
+
+def _maybe_create_general_keep_instance_predicate(cfg: CfgNode) -> Optional[InstancePredicate]:
+    def has_annotations(instance: Instance) -> bool:
+        return "annotations" in instance
+
+    def has_only_crowd_anotations(instance: Instance) -> bool:
+        for ann in instance["annotations"]:
+            if ann.get("is_crowd", 0) == 0:
+                return False
+        return True
+
+    def general_keep_instance_predicate(instance: Instance) -> bool:
+        return has_annotations(instance) and not has_only_crowd_anotations(instance)
+
+    if not cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS:
+        return None
+    return general_keep_instance_predicate
+
+
+def _maybe_create_keypoints_keep_instance_predicate(cfg: CfgNode) -> Optional[InstancePredicate]:
+
+    min_num_keypoints = cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE
+
+    def has_sufficient_num_keypoints(instance: Instance) -> bool:
+        num_kpts = sum(
+            (np.array(ann["keypoints"][2::3]) > 0).sum()
+            for ann in instance["annotations"]
+            if "keypoints" in ann
+        )
+        return num_kpts >= min_num_keypoints
+
+    if cfg.MODEL.KEYPOINT_ON and (min_num_keypoints > 0):
+        return has_sufficient_num_keypoints
+    return None
+
+
+def _maybe_create_mask_keep_instance_predicate(cfg: CfgNode) -> Optional[InstancePredicate]:
+    if not cfg.MODEL.MASK_ON:
+        return None
+
+    def has_mask_annotations(instance: Instance) -> bool:
+        return any("segmentation" in ann for ann in instance["annotations"])
+
+    return has_mask_annotations
+
+
+def _maybe_create_densepose_keep_instance_predicate(cfg: CfgNode) -> Optional[InstancePredicate]:
+    if not cfg.MODEL.DENSEPOSE_ON:
+        return None
+
+    use_masks = cfg.MODEL.ROI_DENSEPOSE_HEAD.COARSE_SEGM_TRAINED_BY_MASKS
+
+    def has_densepose_annotations(instance: Instance) -> bool:
+        for ann in instance["annotations"]:
+            if all(key in ann for key in DENSEPOSE_IUV_KEYS_WITHOUT_MASK) or all(
+                key in ann for key in DENSEPOSE_CSE_KEYS_WITHOUT_MASK
+            ):
+                return True
+            if use_masks and "segmentation" in ann:
+                return True
+        return False
+
+    return has_densepose_annotations
+
+
+def _maybe_create_specific_keep_instance_predicate(cfg: CfgNode) -> Optional[InstancePredicate]:
+    specific_predicate_creators = [
+        _maybe_create_keypoints_keep_instance_predicate,
+        _maybe_create_mask_keep_instance_predicate,
+        _maybe_create_densepose_keep_instance_predicate,
+    ]
+    predicates = [creator(cfg) for creator in specific_predicate_creators]
+    predicates = [p for p in predicates if p is not None]
+    if not predicates:
+        return None
+
+    def combined_predicate(instance: Instance) -> bool:
+        return any(p(instance) for p in predicates)
+
+    return combined_predicate
+
+
+def _get_train_keep_instance_predicate(cfg: CfgNode):
+    general_keep_predicate = _maybe_create_general_keep_instance_predicate(cfg)
+    combined_specific_keep_predicate = _maybe_create_specific_keep_instance_predicate(cfg)
+
+    def combined_general_specific_keep_predicate(instance: Instance) -> bool:
+        return general_keep_predicate(instance) and combined_specific_keep_predicate(instance)
+
+    if (general_keep_predicate is None) and (combined_specific_keep_predicate is None):
+        return None
+    if general_keep_predicate is None:
+        return combined_specific_keep_predicate
+    if combined_specific_keep_predicate is None:
+        return general_keep_predicate
+    return combined_general_specific_keep_predicate
+
+
+def _get_test_keep_instance_predicate(cfg: CfgNode):
+    general_keep_predicate = _maybe_create_general_keep_instance_predicate(cfg)
+    return general_keep_predicate
+
+
+def _maybe_filter_and_map_categories(
+    dataset_name: str, dataset_dicts: List[Instance]
+) -> List[Instance]:
+    meta = MetadataCatalog.get(dataset_name)
+    category_id_map = meta.thing_dataset_id_to_contiguous_id
+    filtered_dataset_dicts = []
+    for dataset_dict in dataset_dicts:
+        anns = []
+        for ann in dataset_dict["annotations"]:
+            cat_id = ann["category_id"]
+            if cat_id not in category_id_map:
+                continue
+            ann["category_id"] = category_id_map[cat_id]
+            anns.append(ann)
+        dataset_dict["annotations"] = anns
+        filtered_dataset_dicts.append(dataset_dict)
+    return filtered_dataset_dicts
+
+
+def _add_category_whitelists_to_metadata(cfg: CfgNode) -> None:
+    for dataset_name, whitelisted_cat_ids in cfg.DATASETS.WHITELISTED_CATEGORIES.items():
+        meta = MetadataCatalog.get(dataset_name)
+        meta.whitelisted_categories = whitelisted_cat_ids
+        logger = logging.getLogger(__name__)
+        logger.info(
+            "Whitelisted categories for dataset {}: {}".format(
+                dataset_name, meta.whitelisted_categories
+            )
+        )
+
+
+def _add_category_maps_to_metadata(cfg: CfgNode) -> None:
+    for dataset_name, category_map in cfg.DATASETS.CATEGORY_MAPS.items():
+        category_map = {
+            int(cat_id_src): int(cat_id_dst) for cat_id_src, cat_id_dst in category_map.items()
+        }
+        meta = MetadataCatalog.get(dataset_name)
+        meta.category_map = category_map
+        logger = logging.getLogger(__name__)
+        logger.info("Category maps for dataset {}: {}".format(dataset_name, meta.category_map))
+
+
+def _add_category_info_to_bootstrapping_metadata(dataset_name: str, dataset_cfg: CfgNode) -> None:
+    meta = MetadataCatalog.get(dataset_name)
+    meta.category_to_class_mapping = get_category_to_class_mapping(dataset_cfg)
+    meta.categories = dataset_cfg.CATEGORIES
+    meta.max_count_per_category = dataset_cfg.MAX_COUNT_PER_CATEGORY
+    logger = logging.getLogger(__name__)
+    logger.info(
+        "Category to class mapping for dataset {}: {}".format(
+            dataset_name, meta.category_to_class_mapping
+        )
+    )
+
+
+def _maybe_add_class_to_mesh_name_map_to_metadata(dataset_names: List[str], cfg: CfgNode) -> None:
+    for dataset_name in dataset_names:
+        meta = MetadataCatalog.get(dataset_name)
+        if not hasattr(meta, "class_to_mesh_name"):
+            meta.class_to_mesh_name = get_class_to_mesh_name_mapping(cfg)
+
+
+def _merge_categories(dataset_names: Collection[str]) -> _MergedCategoriesT:
+    merged_categories = defaultdict(list)
+    category_names = {}
+    for dataset_name in dataset_names:
+        meta = MetadataCatalog.get(dataset_name)
+        whitelisted_categories = meta.get("whitelisted_categories")
+        category_map = meta.get("category_map", {})
+        cat_ids = (
+            whitelisted_categories if whitelisted_categories is not None else meta.categories.keys()
+        )
+        for cat_id in cat_ids:
+            cat_name = meta.categories[cat_id]
+            cat_id_mapped = category_map.get(cat_id, cat_id)
+            if cat_id_mapped == cat_id or cat_id_mapped in cat_ids:
+                category_names[cat_id] = cat_name
+            else:
+                category_names[cat_id] = str(cat_id_mapped)
+            # assign temporary mapped category name, this name can be changed
+            # during the second pass, since mapped ID can correspond to a category
+            # from a different dataset
+            cat_name_mapped = meta.categories[cat_id_mapped]
+            merged_categories[cat_id_mapped].append(
+                _DatasetCategory(
+                    id=cat_id,
+                    name=cat_name,
+                    mapped_id=cat_id_mapped,
+                    mapped_name=cat_name_mapped,
+                    dataset_name=dataset_name,
+                )
+            )
+    # second pass to assign proper mapped category names
+    for cat_id, categories in merged_categories.items():
+        for cat in categories:
+            if cat_id in category_names and cat.mapped_name != category_names[cat_id]:
+                cat.mapped_name = category_names[cat_id]
+
+    return merged_categories
+
+
+def _warn_if_merged_different_categories(merged_categories: _MergedCategoriesT) -> None:
+    logger = logging.getLogger(__name__)
+    for cat_id in merged_categories:
+        merged_categories_i = merged_categories[cat_id]
+        first_cat_name = merged_categories_i[0].name
+        if len(merged_categories_i) > 1 and not all(
+            cat.name == first_cat_name for cat in merged_categories_i[1:]
+        ):
+            cat_summary_str = ", ".join(
+                [f"{cat.id} ({cat.name}) from {cat.dataset_name}" for cat in merged_categories_i]
+            )
+            logger.warning(
+                f"Merged category {cat_id} corresponds to the following categories: "
+                f"{cat_summary_str}"
+            )
+
+
+def combine_detection_dataset_dicts(
+    dataset_names: Collection[str],
+    keep_instance_predicate: Optional[InstancePredicate] = None,
+    proposal_files: Optional[Collection[str]] = None,
+) -> List[Instance]:
+    """
+    Load and prepare dataset dicts for training / testing
+
+    Args:
+        dataset_names (Collection[str]): a list of dataset names
+        keep_instance_predicate (Callable: Dict[str, Any] -> bool): predicate
+            applied to instance dicts which defines whether to keep the instance
+        proposal_files (Collection[str]): if given, a list of object proposal files
+            that match each dataset in `dataset_names`.
+    """
+    assert len(dataset_names)
+    if proposal_files is None:
+        proposal_files = [None] * len(dataset_names)
+    assert len(dataset_names) == len(proposal_files)
+    # load datasets and metadata
+    dataset_name_to_dicts = {}
+    for dataset_name in dataset_names:
+        dataset_name_to_dicts[dataset_name] = DatasetCatalog.get(dataset_name)
+        assert len(dataset_name_to_dicts), f"Dataset '{dataset_name}' is empty!"
+    # merge categories, requires category metadata to be loaded
+    # cat_id -> [(orig_cat_id, cat_name, dataset_name)]
+    merged_categories = _merge_categories(dataset_names)
+    _warn_if_merged_different_categories(merged_categories)
+    merged_category_names = [
+        merged_categories[cat_id][0].mapped_name for cat_id in sorted(merged_categories)
+    ]
+    # map to contiguous category IDs
+    _add_category_id_to_contiguous_id_maps_to_metadata(merged_categories)
+    # load annotations and dataset metadata
+    for dataset_name, proposal_file in zip(dataset_names, proposal_files):
+        dataset_dicts = dataset_name_to_dicts[dataset_name]
+        assert len(dataset_dicts), f"Dataset '{dataset_name}' is empty!"
+        if proposal_file is not None:
+            dataset_dicts = load_proposals_into_dataset(dataset_dicts, proposal_file)
+        dataset_dicts = _maybe_filter_and_map_categories(dataset_name, dataset_dicts)
+        print_instances_class_histogram(dataset_dicts, merged_category_names)
+        dataset_name_to_dicts[dataset_name] = dataset_dicts
+
+    if keep_instance_predicate is not None:
+        all_datasets_dicts_plain = [
+            d
+            for d in itertools.chain.from_iterable(dataset_name_to_dicts.values())
+            if keep_instance_predicate(d)
+        ]
+    else:
+        all_datasets_dicts_plain = list(
+            itertools.chain.from_iterable(dataset_name_to_dicts.values())
+        )
+    return all_datasets_dicts_plain
+
+
+def build_detection_train_loader(cfg: CfgNode, mapper=None):
+    """
+    A data loader is created in a way similar to that of Detectron2.
+    The main differences are:
+     - it allows to combine datasets with different but compatible object category sets
+
+    The data loader is created by the following steps:
+    1. Use the dataset names in config to query :class:`DatasetCatalog`, and obtain a list of dicts.
+    2. Start workers to work on the dicts. Each worker will:
+        * Map each metadata dict into another format to be consumed by the model.
+        * Batch them by simply putting dicts into a list.
+    The batched ``list[mapped_dict]`` is what this dataloader will return.
+
+    Args:
+        cfg (CfgNode): the config
+        mapper (callable): a callable which takes a sample (dict) from dataset and
+            returns the format to be consumed by the model.
+            By default it will be `DatasetMapper(cfg, True)`.
+
+    Returns:
+        an infinite iterator of training data
+    """
+
+    _add_category_whitelists_to_metadata(cfg)
+    _add_category_maps_to_metadata(cfg)
+    _maybe_add_class_to_mesh_name_map_to_metadata(cfg.DATASETS.TRAIN, cfg)
+    dataset_dicts = combine_detection_dataset_dicts(
+        cfg.DATASETS.TRAIN,
+        keep_instance_predicate=_get_train_keep_instance_predicate(cfg),
+        proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None,
+    )
+    if mapper is None:
+        mapper = DatasetMapper(cfg, True)
+    return d2_build_detection_train_loader(cfg, dataset=dataset_dicts, mapper=mapper)
+
+
+def build_detection_test_loader(cfg, dataset_name, mapper=None):
+    """
+    Similar to `build_detection_train_loader`.
+    But this function uses the given `dataset_name` argument (instead of the names in cfg),
+    and uses batch size 1.
+
+    Args:
+        cfg: a detectron2 CfgNode
+        dataset_name (str): a name of the dataset that's available in the DatasetCatalog
+        mapper (callable): a callable which takes a sample (dict) from dataset
+            and returns the format to be consumed by the model.
+            By default it will be `DatasetMapper(cfg, False)`.
+
+    Returns:
+        DataLoader: a torch DataLoader, that loads the given detection
+            dataset, with test-time transformation and batching.
+    """
+    _add_category_whitelists_to_metadata(cfg)
+    _add_category_maps_to_metadata(cfg)
+    _maybe_add_class_to_mesh_name_map_to_metadata([dataset_name], cfg)
+    dataset_dicts = combine_detection_dataset_dicts(
+        [dataset_name],
+        keep_instance_predicate=_get_test_keep_instance_predicate(cfg),
+        proposal_files=[
+            cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(dataset_name)]
+        ]
+        if cfg.MODEL.LOAD_PROPOSALS
+        else None,
+    )
+    sampler = None
+    if not cfg.DENSEPOSE_EVALUATION.DISTRIBUTED_INFERENCE:
+        sampler = torch.utils.data.SequentialSampler(dataset_dicts)
+    if mapper is None:
+        mapper = DatasetMapper(cfg, False)
+    return d2_build_detection_test_loader(
+        dataset_dicts, mapper=mapper, num_workers=cfg.DATALOADER.NUM_WORKERS, sampler=sampler
+    )
+
+
+def build_frame_selector(cfg: CfgNode):
+    strategy = FrameSelectionStrategy(cfg.STRATEGY)
+    if strategy == FrameSelectionStrategy.RANDOM_K:
+        frame_selector = RandomKFramesSelector(cfg.NUM_IMAGES)
+    elif strategy == FrameSelectionStrategy.FIRST_K:
+        frame_selector = FirstKFramesSelector(cfg.NUM_IMAGES)
+    elif strategy == FrameSelectionStrategy.LAST_K:
+        frame_selector = LastKFramesSelector(cfg.NUM_IMAGES)
+    elif strategy == FrameSelectionStrategy.ALL:
+        frame_selector = None
+    # pyre-fixme[61]: `frame_selector` may not be initialized here.
+    return frame_selector
+
+
+def build_transform(cfg: CfgNode, data_type: str):
+    if cfg.TYPE == "resize":
+        if data_type == "image":
+            return ImageResizeTransform(cfg.MIN_SIZE, cfg.MAX_SIZE)
+    raise ValueError(f"Unknown transform {cfg.TYPE} for data type {data_type}")
+
+
+def build_combined_loader(cfg: CfgNode, loaders: Collection[Loader], ratios: Sequence[float]):
+    images_per_worker = _compute_num_images_per_worker(cfg)
+    return CombinedDataLoader(loaders, images_per_worker, ratios)
+
+
+def build_bootstrap_dataset(dataset_name: str, cfg: CfgNode) -> Sequence[torch.Tensor]:
+    """
+    Build dataset that provides data to bootstrap on
+
+    Args:
+        dataset_name (str): Name of the dataset, needs to have associated metadata
+            to load the data
+        cfg (CfgNode): bootstrapping config
+    Returns:
+        Sequence[Tensor] - dataset that provides image batches, Tensors of size
+            [N, C, H, W] of type float32
+    """
+    logger = logging.getLogger(__name__)
+    _add_category_info_to_bootstrapping_metadata(dataset_name, cfg)
+    meta = MetadataCatalog.get(dataset_name)
+    factory = BootstrapDatasetFactoryCatalog.get(meta.dataset_type)
+    dataset = None
+    if factory is not None:
+        dataset = factory(meta, cfg)
+    if dataset is None:
+        logger.warning(f"Failed to create dataset {dataset_name} of type {meta.dataset_type}")
+    return dataset
+
+
+def build_data_sampler(cfg: CfgNode, sampler_cfg: CfgNode, embedder: Optional[torch.nn.Module]):
+    if sampler_cfg.TYPE == "densepose_uniform":
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseUniformSampler(count_per_class=sampler_cfg.COUNT_PER_CLASS),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+    elif sampler_cfg.TYPE == "densepose_UV_confidence":
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseConfidenceBasedSampler(
+                confidence_channel="sigma_2",
+                count_per_class=sampler_cfg.COUNT_PER_CLASS,
+                search_proportion=0.5,
+            ),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+    elif sampler_cfg.TYPE == "densepose_fine_segm_confidence":
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseConfidenceBasedSampler(
+                confidence_channel="fine_segm_confidence",
+                count_per_class=sampler_cfg.COUNT_PER_CLASS,
+                search_proportion=0.5,
+            ),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+    elif sampler_cfg.TYPE == "densepose_coarse_segm_confidence":
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseConfidenceBasedSampler(
+                confidence_channel="coarse_segm_confidence",
+                count_per_class=sampler_cfg.COUNT_PER_CLASS,
+                search_proportion=0.5,
+            ),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+    elif sampler_cfg.TYPE == "densepose_cse_uniform":
+        assert embedder is not None
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseCSEUniformSampler(
+                cfg=cfg,
+                use_gt_categories=sampler_cfg.USE_GROUND_TRUTH_CATEGORIES,
+                embedder=embedder,
+                count_per_class=sampler_cfg.COUNT_PER_CLASS,
+            ),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+    elif sampler_cfg.TYPE == "densepose_cse_coarse_segm_confidence":
+        assert embedder is not None
+        data_sampler = PredictionToGroundTruthSampler()
+        # transform densepose pred -> gt
+        data_sampler.register_sampler(
+            "pred_densepose",
+            "gt_densepose",
+            DensePoseCSEConfidenceBasedSampler(
+                cfg=cfg,
+                use_gt_categories=sampler_cfg.USE_GROUND_TRUTH_CATEGORIES,
+                embedder=embedder,
+                confidence_channel="coarse_segm_confidence",
+                count_per_class=sampler_cfg.COUNT_PER_CLASS,
+                search_proportion=0.5,
+            ),
+        )
+        data_sampler.register_sampler("pred_densepose", "gt_masks", MaskFromDensePoseSampler())
+        return data_sampler
+
+    raise ValueError(f"Unknown data sampler type {sampler_cfg.TYPE}")
+
+
+def build_data_filter(cfg: CfgNode):
+    if cfg.TYPE == "detection_score":
+        min_score = cfg.MIN_VALUE
+        return ScoreBasedFilter(min_score=min_score)
+    raise ValueError(f"Unknown data filter type {cfg.TYPE}")
+
+
+def build_inference_based_loader(
+    cfg: CfgNode,
+    dataset_cfg: CfgNode,
+    model: torch.nn.Module,
+    embedder: Optional[torch.nn.Module] = None,
+) -> InferenceBasedLoader:
+    """
+    Constructs data loader based on inference results of a model.
+    """
+    dataset = build_bootstrap_dataset(dataset_cfg.DATASET, dataset_cfg.IMAGE_LOADER)
+    meta = MetadataCatalog.get(dataset_cfg.DATASET)
+    training_sampler = TrainingSampler(len(dataset))
+    data_loader = torch.utils.data.DataLoader(
+        dataset,  # pyre-ignore[6]
+        batch_size=dataset_cfg.IMAGE_LOADER.BATCH_SIZE,
+        sampler=training_sampler,
+        num_workers=dataset_cfg.IMAGE_LOADER.NUM_WORKERS,
+        collate_fn=trivial_batch_collator,
+        worker_init_fn=worker_init_reset_seed,
+    )
+    return InferenceBasedLoader(
+        model,
+        data_loader=data_loader,
+        data_sampler=build_data_sampler(cfg, dataset_cfg.DATA_SAMPLER, embedder),
+        data_filter=build_data_filter(dataset_cfg.FILTER),
+        shuffle=True,
+        batch_size=dataset_cfg.INFERENCE.OUTPUT_BATCH_SIZE,
+        inference_batch_size=dataset_cfg.INFERENCE.INPUT_BATCH_SIZE,
+        category_to_class_mapping=meta.category_to_class_mapping,
+    )
+
+
+def has_inference_based_loaders(cfg: CfgNode) -> bool:
+    """
+    Returns True, if at least one inferense-based loader must
+    be instantiated for training
+    """
+    return len(cfg.BOOTSTRAP_DATASETS) > 0
+
+
+def build_inference_based_loaders(
+    cfg: CfgNode, model: torch.nn.Module
+) -> Tuple[List[InferenceBasedLoader], List[float]]:
+    loaders = []
+    ratios = []
+    embedder = build_densepose_embedder(cfg).to(device=model.device)  # pyre-ignore[16]
+    for dataset_spec in cfg.BOOTSTRAP_DATASETS:
+        dataset_cfg = get_bootstrap_dataset_config().clone()
+        dataset_cfg.merge_from_other_cfg(CfgNode(dataset_spec))
+        loader = build_inference_based_loader(cfg, dataset_cfg, model, embedder)
+        loaders.append(loader)
+        ratios.append(dataset_cfg.RATIO)
+    return loaders, ratios
+
+
+def build_video_list_dataset(meta: Metadata, cfg: CfgNode):
+    video_list_fpath = meta.video_list_fpath
+    video_base_path = meta.video_base_path
+    category = meta.category
+    if cfg.TYPE == "video_keyframe":
+        frame_selector = build_frame_selector(cfg.SELECT)
+        transform = build_transform(cfg.TRANSFORM, data_type="image")
+        video_list = video_list_from_file(video_list_fpath, video_base_path)
+        keyframe_helper_fpath = getattr(cfg, "KEYFRAME_HELPER", None)
+        return VideoKeyframeDataset(
+            video_list, category, frame_selector, transform, keyframe_helper_fpath
+        )
+
+
+class _BootstrapDatasetFactoryCatalog(UserDict):
+    """
+    A global dictionary that stores information about bootstrapped datasets creation functions
+    from metadata and config, for diverse DatasetType
+    """
+
+    def register(self, dataset_type: DatasetType, factory: Callable[[Metadata, CfgNode], Dataset]):
+        """
+        Args:
+            dataset_type (DatasetType): a DatasetType e.g. DatasetType.VIDEO_LIST
+            factory (Callable[Metadata, CfgNode]): a callable which takes Metadata and cfg
+            arguments and returns a dataset object.
+        """
+        assert dataset_type not in self, "Dataset '{}' is already registered!".format(dataset_type)
+        self[dataset_type] = factory
+
+
+BootstrapDatasetFactoryCatalog = _BootstrapDatasetFactoryCatalog()
+BootstrapDatasetFactoryCatalog.register(DatasetType.VIDEO_LIST, build_video_list_dataset)

3rdparty/densepose/data/combined_loader.py

@@ -0,0 +1,44 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+from collections import deque
+from typing import Any, Collection, Deque, Iterable, Iterator, List, Sequence
+
+Loader = Iterable[Any]
+
+
+def _pooled_next(iterator: Iterator[Any], pool: Deque[Any]):
+    if not pool:
+        pool.extend(next(iterator))
+    return pool.popleft()
+
+
+class CombinedDataLoader:
+    """
+    Combines data loaders using the provided sampling ratios
+    """
+
+    BATCH_COUNT = 100
+
+    def __init__(self, loaders: Collection[Loader], batch_size: int, ratios: Sequence[float]):
+        self.loaders = loaders
+        self.batch_size = batch_size
+        self.ratios = ratios
+
+    def __iter__(self) -> Iterator[List[Any]]:
+        iters = [iter(loader) for loader in self.loaders]
+        indices = []
+        pool = [deque()] * len(iters)
+        # infinite iterator, as in D2
+        while True:
+            if not indices:
+                # just a buffer of indices, its size doesn't matter
+                # as long as it's a multiple of batch_size
+                k = self.batch_size * self.BATCH_COUNT
+                indices = random.choices(range(len(self.loaders)), self.ratios, k=k)
+            try:
+                batch = [_pooled_next(iters[i], pool[i]) for i in indices[: self.batch_size]]
+            except StopIteration:
+                break
+            indices = indices[self.batch_size :]
+            yield batch

3rdparty/densepose/data/dataset_mapper.py

@@ -0,0 +1,168 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import copy
+import logging
+from typing import Any, Dict, List, Tuple
+import torch
+
+from detectron2.data import MetadataCatalog
+from detectron2.data import detection_utils as utils
+from detectron2.data import transforms as T
+from detectron2.layers import ROIAlign
+from detectron2.structures import BoxMode
+from detectron2.utils.file_io import PathManager
+
+from densepose.structures import DensePoseDataRelative, DensePoseList, DensePoseTransformData
+
+
+def build_augmentation(cfg, is_train):
+    logger = logging.getLogger(__name__)
+    result = utils.build_augmentation(cfg, is_train)
+    if is_train:
+        random_rotation = T.RandomRotation(
+            cfg.INPUT.ROTATION_ANGLES, expand=False, sample_style="choice"
+        )
+        result.append(random_rotation)
+        logger.info("DensePose-specific augmentation used in training: " + str(random_rotation))
+    return result
+
+
+class DatasetMapper:
+    """
+    A customized version of `detectron2.data.DatasetMapper`
+    """
+
+    def __init__(self, cfg, is_train=True):
+        self.augmentation = build_augmentation(cfg, is_train)
+
+        # fmt: off
+        self.img_format     = cfg.INPUT.FORMAT
+        self.mask_on        = (
+            cfg.MODEL.MASK_ON or (
+                cfg.MODEL.DENSEPOSE_ON
+                and cfg.MODEL.ROI_DENSEPOSE_HEAD.COARSE_SEGM_TRAINED_BY_MASKS)
+        )
+        self.keypoint_on    = cfg.MODEL.KEYPOINT_ON
+        self.densepose_on   = cfg.MODEL.DENSEPOSE_ON
+        assert not cfg.MODEL.LOAD_PROPOSALS, "not supported yet"
+        # fmt: on
+        if self.keypoint_on and is_train:
+            # Flip only makes sense in training
+            self.keypoint_hflip_indices = utils.create_keypoint_hflip_indices(cfg.DATASETS.TRAIN)
+        else:
+            self.keypoint_hflip_indices = None
+
+        if self.densepose_on:
+            densepose_transform_srcs = [
+                MetadataCatalog.get(ds).densepose_transform_src
+                for ds in cfg.DATASETS.TRAIN + cfg.DATASETS.TEST
+            ]
+            assert len(densepose_transform_srcs) > 0
+            # TODO: check that DensePose transformation data is the same for
+            # all the datasets. Otherwise one would have to pass DB ID with
+            # each entry to select proper transformation data. For now, since
+            # all DensePose annotated data uses the same data semantics, we
+            # omit this check.
+            densepose_transform_data_fpath = PathManager.get_local_path(densepose_transform_srcs[0])
+            self.densepose_transform_data = DensePoseTransformData.load(
+                densepose_transform_data_fpath
+            )
+
+        self.is_train = is_train
+
+    def __call__(self, dataset_dict):
+        """
+        Args:
+            dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
+
+        Returns:
+            dict: a format that builtin models in detectron2 accept
+        """
+        dataset_dict = copy.deepcopy(dataset_dict)  # it will be modified by code below
+        image = utils.read_image(dataset_dict["file_name"], format=self.img_format)
+        utils.check_image_size(dataset_dict, image)
+
+        image, transforms = T.apply_transform_gens(self.augmentation, image)
+        image_shape = image.shape[:2]  # h, w
+        dataset_dict["image"] = torch.as_tensor(image.transpose(2, 0, 1).astype("float32"))
+
+        if not self.is_train:
+            dataset_dict.pop("annotations", None)
+            return dataset_dict
+
+        for anno in dataset_dict["annotations"]:
+            if not self.mask_on:
+                anno.pop("segmentation", None)
+            if not self.keypoint_on:
+                anno.pop("keypoints", None)
+
+        # USER: Implement additional transformations if you have other types of data
+        # USER: Don't call transpose_densepose if you don't need
+        annos = [
+            self._transform_densepose(
+                utils.transform_instance_annotations(
+                    obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices
+                ),
+                transforms,
+            )
+            for obj in dataset_dict.pop("annotations")
+            if obj.get("iscrowd", 0) == 0
+        ]
+
+        if self.mask_on:
+            self._add_densepose_masks_as_segmentation(annos, image_shape)
+
+        instances = utils.annotations_to_instances(annos, image_shape, mask_format="bitmask")
+        densepose_annotations = [obj.get("densepose") for obj in annos]
+        if densepose_annotations and not all(v is None for v in densepose_annotations):
+            instances.gt_densepose = DensePoseList(
+                densepose_annotations, instances.gt_boxes, image_shape
+            )
+
+        dataset_dict["instances"] = instances[instances.gt_boxes.nonempty()]
+        return dataset_dict
+
+    def _transform_densepose(self, annotation, transforms):
+        if not self.densepose_on:
+            return annotation
+
+        # Handle densepose annotations
+        is_valid, reason_not_valid = DensePoseDataRelative.validate_annotation(annotation)
+        if is_valid:
+            densepose_data = DensePoseDataRelative(annotation, cleanup=True)
+            densepose_data.apply_transform(transforms, self.densepose_transform_data)
+            annotation["densepose"] = densepose_data
+        else:
+            # logger = logging.getLogger(__name__)
+            # logger.debug("Could not load DensePose annotation: {}".format(reason_not_valid))
+            DensePoseDataRelative.cleanup_annotation(annotation)
+            # NOTE: annotations for certain instances may be unavailable.
+            # 'None' is accepted by the DensePostList data structure.
+            annotation["densepose"] = None
+        return annotation
+
+    def _add_densepose_masks_as_segmentation(
+        self, annotations: List[Dict[str, Any]], image_shape_hw: Tuple[int, int]
+    ):
+        for obj in annotations:
+            if ("densepose" not in obj) or ("segmentation" in obj):
+                continue
+            # DP segmentation: torch.Tensor [S, S] of float32, S=256
+            segm_dp = torch.zeros_like(obj["densepose"].segm)
+            segm_dp[obj["densepose"].segm > 0] = 1
+            segm_h, segm_w = segm_dp.shape
+            bbox_segm_dp = torch.tensor((0, 0, segm_h - 1, segm_w - 1), dtype=torch.float32)
+            # image bbox
+            x0, y0, x1, y1 = (
+                v.item() for v in BoxMode.convert(obj["bbox"], obj["bbox_mode"], BoxMode.XYXY_ABS)
+            )
+            segm_aligned = (
+                ROIAlign((y1 - y0, x1 - x0), 1.0, 0, aligned=True)
+                .forward(segm_dp.view(1, 1, *segm_dp.shape), bbox_segm_dp)
+                .squeeze()
+            )
+            image_mask = torch.zeros(*image_shape_hw, dtype=torch.float32)
+            image_mask[y0:y1, x0:x1] = segm_aligned
+            # segmentation for BitMask: np.array [H, W] of bool
+            obj["segmentation"] = image_mask >= 0.5

3rdparty/densepose/data/datasets/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from . import builtin  # ensure the builtin datasets are registered
+
+__all__ = [k for k in globals().keys() if "builtin" not in k and not k.startswith("_")]

3rdparty/densepose/data/datasets/builtin.py

@@ -0,0 +1,16 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from .chimpnsee import register_dataset as register_chimpnsee_dataset
+from .coco import BASE_DATASETS as BASE_COCO_DATASETS
+from .coco import DATASETS as COCO_DATASETS
+from .coco import register_datasets as register_coco_datasets
+from .lvis import DATASETS as LVIS_DATASETS
+from .lvis import register_datasets as register_lvis_datasets
+
+DEFAULT_DATASETS_ROOT = "datasets"
+
+
+register_coco_datasets(COCO_DATASETS, DEFAULT_DATASETS_ROOT)
+register_coco_datasets(BASE_COCO_DATASETS, DEFAULT_DATASETS_ROOT)
+register_lvis_datasets(LVIS_DATASETS, DEFAULT_DATASETS_ROOT)
+
+register_chimpnsee_dataset(DEFAULT_DATASETS_ROOT)  # pyre-ignore[19]

3rdparty/densepose/data/datasets/chimpnsee.py

@@ -0,0 +1,29 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Optional
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+
+from ..utils import maybe_prepend_base_path
+from .dataset_type import DatasetType
+
+CHIMPNSEE_DATASET_NAME = "chimpnsee"
+
+
+def register_dataset(datasets_root: Optional[str] = None) -> None:
+    def empty_load_callback():
+        pass
+
+    video_list_fpath = maybe_prepend_base_path(
+        datasets_root,
+        "chimpnsee/cdna.eva.mpg.de/video_list.txt",
+    )
+    video_base_path = maybe_prepend_base_path(datasets_root, "chimpnsee/cdna.eva.mpg.de")
+
+    DatasetCatalog.register(CHIMPNSEE_DATASET_NAME, empty_load_callback)
+    MetadataCatalog.get(CHIMPNSEE_DATASET_NAME).set(
+        dataset_type=DatasetType.VIDEO_LIST,
+        video_list_fpath=video_list_fpath,
+        video_base_path=video_base_path,
+        category="chimpanzee",
+    )

3rdparty/densepose/data/datasets/coco.py

@@ -0,0 +1,432 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import contextlib
+import io
+import logging
+import os
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import Any, Dict, Iterable, List, Optional
+from fvcore.common.timer import Timer
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+from detectron2.structures import BoxMode
+from detectron2.utils.file_io import PathManager
+
+from ..utils import maybe_prepend_base_path
+
+DENSEPOSE_MASK_KEY = "dp_masks"
+DENSEPOSE_IUV_KEYS_WITHOUT_MASK = ["dp_x", "dp_y", "dp_I", "dp_U", "dp_V"]
+DENSEPOSE_CSE_KEYS_WITHOUT_MASK = ["dp_x", "dp_y", "dp_vertex", "ref_model"]
+DENSEPOSE_ALL_POSSIBLE_KEYS = set(
+    DENSEPOSE_IUV_KEYS_WITHOUT_MASK + DENSEPOSE_CSE_KEYS_WITHOUT_MASK + [DENSEPOSE_MASK_KEY]
+)
+DENSEPOSE_METADATA_URL_PREFIX = "https://dl.fbaipublicfiles.com/densepose/data/"
+
+
+@dataclass
+class CocoDatasetInfo:
+    name: str
+    images_root: str
+    annotations_fpath: str
+
+
+DATASETS = [
+    CocoDatasetInfo(
+        name="densepose_coco_2014_train",
+        images_root="coco/train2014",
+        annotations_fpath="coco/annotations/densepose_train2014.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_minival",
+        images_root="coco/val2014",
+        annotations_fpath="coco/annotations/densepose_minival2014.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_minival_100",
+        images_root="coco/val2014",
+        annotations_fpath="coco/annotations/densepose_minival2014_100.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_valminusminival",
+        images_root="coco/val2014",
+        annotations_fpath="coco/annotations/densepose_valminusminival2014.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_train_cse",
+        images_root="coco/train2014",
+        annotations_fpath="coco_cse/densepose_train2014_cse.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_minival_cse",
+        images_root="coco/val2014",
+        annotations_fpath="coco_cse/densepose_minival2014_cse.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_minival_100_cse",
+        images_root="coco/val2014",
+        annotations_fpath="coco_cse/densepose_minival2014_100_cse.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_coco_2014_valminusminival_cse",
+        images_root="coco/val2014",
+        annotations_fpath="coco_cse/densepose_valminusminival2014_cse.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_chimps",
+        images_root="densepose_chimps/images",
+        annotations_fpath="densepose_chimps/densepose_chimps_densepose.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_chimps_cse_train",
+        images_root="densepose_chimps/images",
+        annotations_fpath="densepose_chimps/densepose_chimps_cse_train.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_chimps_cse_val",
+        images_root="densepose_chimps/images",
+        annotations_fpath="densepose_chimps/densepose_chimps_cse_val.json",
+    ),
+    CocoDatasetInfo(
+        name="posetrack2017_train",
+        images_root="posetrack2017/posetrack_data_2017",
+        annotations_fpath="posetrack2017/densepose_posetrack_train2017.json",
+    ),
+    CocoDatasetInfo(
+        name="posetrack2017_val",
+        images_root="posetrack2017/posetrack_data_2017",
+        annotations_fpath="posetrack2017/densepose_posetrack_val2017.json",
+    ),
+    CocoDatasetInfo(
+        name="lvis_v05_train",
+        images_root="coco/train2017",
+        annotations_fpath="lvis/lvis_v0.5_plus_dp_train.json",
+    ),
+    CocoDatasetInfo(
+        name="lvis_v05_val",
+        images_root="coco/val2017",
+        annotations_fpath="lvis/lvis_v0.5_plus_dp_val.json",
+    ),
+]
+
+
+BASE_DATASETS = [
+    CocoDatasetInfo(
+        name="base_coco_2017_train",
+        images_root="coco/train2017",
+        annotations_fpath="coco/annotations/instances_train2017.json",
+    ),
+    CocoDatasetInfo(
+        name="base_coco_2017_val",
+        images_root="coco/val2017",
+        annotations_fpath="coco/annotations/instances_val2017.json",
+    ),
+    CocoDatasetInfo(
+        name="base_coco_2017_val_100",
+        images_root="coco/val2017",
+        annotations_fpath="coco/annotations/instances_val2017_100.json",
+    ),
+]
+
+
+def get_metadata(base_path: Optional[str]) -> Dict[str, Any]:
+    """
+    Returns metadata associated with COCO DensePose datasets
+
+    Args:
+    base_path: Optional[str]
+        Base path used to load metadata from
+
+    Returns:
+    Dict[str, Any]
+        Metadata in the form of a dictionary
+    """
+    meta = {
+        "densepose_transform_src": maybe_prepend_base_path(base_path, "UV_symmetry_transforms.mat"),
+        "densepose_smpl_subdiv": maybe_prepend_base_path(base_path, "SMPL_subdiv.mat"),
+        "densepose_smpl_subdiv_transform": maybe_prepend_base_path(
+            base_path,
+            "SMPL_SUBDIV_TRANSFORM.mat",
+        ),
+    }
+    return meta
+
+
+def _load_coco_annotations(json_file: str):
+    """
+    Load COCO annotations from a JSON file
+
+    Args:
+        json_file: str
+            Path to the file to load annotations from
+    Returns:
+        Instance of `pycocotools.coco.COCO` that provides access to annotations
+        data
+    """
+    from pycocotools.coco import COCO
+
+    logger = logging.getLogger(__name__)
+    timer = Timer()
+    with contextlib.redirect_stdout(io.StringIO()):
+        coco_api = COCO(json_file)
+    if timer.seconds() > 1:
+        logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds()))
+    return coco_api
+
+
+def _add_categories_metadata(dataset_name: str, categories: List[Dict[str, Any]]):
+    meta = MetadataCatalog.get(dataset_name)
+    meta.categories = {c["id"]: c["name"] for c in categories}
+    logger = logging.getLogger(__name__)
+    logger.info("Dataset {} categories: {}".format(dataset_name, meta.categories))
+
+
+def _verify_annotations_have_unique_ids(json_file: str, anns: List[List[Dict[str, Any]]]):
+    if "minival" in json_file:
+        # Skip validation on COCO2014 valminusminival and minival annotations
+        # The ratio of buggy annotations there is tiny and does not affect accuracy
+        # Therefore we explicitly white-list them
+        return
+    ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image]
+    assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique!".format(
+        json_file
+    )
+
+
+def _maybe_add_bbox(obj: Dict[str, Any], ann_dict: Dict[str, Any]):
+    if "bbox" not in ann_dict:
+        return
+    obj["bbox"] = ann_dict["bbox"]
+    obj["bbox_mode"] = BoxMode.XYWH_ABS
+
+
+def _maybe_add_segm(obj: Dict[str, Any], ann_dict: Dict[str, Any]):
+    if "segmentation" not in ann_dict:
+        return
+    segm = ann_dict["segmentation"]
+    if not isinstance(segm, dict):
+        # filter out invalid polygons (< 3 points)
+        segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6]
+        if len(segm) == 0:
+            return
+    obj["segmentation"] = segm
+
+
+def _maybe_add_keypoints(obj: Dict[str, Any], ann_dict: Dict[str, Any]):
+    if "keypoints" not in ann_dict:
+        return
+    keypts = ann_dict["keypoints"]  # list[int]
+    for idx, v in enumerate(keypts):
+        if idx % 3 != 2:
+            # COCO's segmentation coordinates are floating points in [0, H or W],
+            # but keypoint coordinates are integers in [0, H-1 or W-1]
+            # Therefore we assume the coordinates are "pixel indices" and
+            # add 0.5 to convert to floating point coordinates.
+            keypts[idx] = v + 0.5
+    obj["keypoints"] = keypts
+
+
+def _maybe_add_densepose(obj: Dict[str, Any], ann_dict: Dict[str, Any]):
+    for key in DENSEPOSE_ALL_POSSIBLE_KEYS:
+        if key in ann_dict:
+            obj[key] = ann_dict[key]
+
+
+def _combine_images_with_annotations(
+    dataset_name: str,
+    image_root: str,
+    img_datas: Iterable[Dict[str, Any]],
+    ann_datas: Iterable[Iterable[Dict[str, Any]]],
+):
+
+    ann_keys = ["iscrowd", "category_id"]
+    dataset_dicts = []
+    contains_video_frame_info = False
+
+    for img_dict, ann_dicts in zip(img_datas, ann_datas):
+        record = {}
+        record["file_name"] = os.path.join(image_root, img_dict["file_name"])
+        record["height"] = img_dict["height"]
+        record["width"] = img_dict["width"]
+        record["image_id"] = img_dict["id"]
+        record["dataset"] = dataset_name
+        if "frame_id" in img_dict:
+            record["frame_id"] = img_dict["frame_id"]
+            record["video_id"] = img_dict.get("vid_id", None)
+            contains_video_frame_info = True
+        objs = []
+        for ann_dict in ann_dicts:
+            assert ann_dict["image_id"] == record["image_id"]
+            assert ann_dict.get("ignore", 0) == 0
+            obj = {key: ann_dict[key] for key in ann_keys if key in ann_dict}
+            _maybe_add_bbox(obj, ann_dict)
+            _maybe_add_segm(obj, ann_dict)
+            _maybe_add_keypoints(obj, ann_dict)
+            _maybe_add_densepose(obj, ann_dict)
+            objs.append(obj)
+        record["annotations"] = objs
+        dataset_dicts.append(record)
+    if contains_video_frame_info:
+        create_video_frame_mapping(dataset_name, dataset_dicts)
+    return dataset_dicts
+
+
+def get_contiguous_id_to_category_id_map(metadata):
+    cat_id_2_cont_id = metadata.thing_dataset_id_to_contiguous_id
+    cont_id_2_cat_id = {}
+    for cat_id, cont_id in cat_id_2_cont_id.items():
+        if cont_id in cont_id_2_cat_id:
+            continue
+        cont_id_2_cat_id[cont_id] = cat_id
+    return cont_id_2_cat_id
+
+
+def maybe_filter_categories_cocoapi(dataset_name, coco_api):
+    meta = MetadataCatalog.get(dataset_name)
+    cont_id_2_cat_id = get_contiguous_id_to_category_id_map(meta)
+    cat_id_2_cont_id = meta.thing_dataset_id_to_contiguous_id
+    # filter categories
+    cats = []
+    for cat in coco_api.dataset["categories"]:
+        cat_id = cat["id"]
+        if cat_id not in cat_id_2_cont_id:
+            continue
+        cont_id = cat_id_2_cont_id[cat_id]
+        if (cont_id in cont_id_2_cat_id) and (cont_id_2_cat_id[cont_id] == cat_id):
+            cats.append(cat)
+    coco_api.dataset["categories"] = cats
+    # filter annotations, if multiple categories are mapped to a single
+    # contiguous ID, use only one category ID and map all annotations to that category ID
+    anns = []
+    for ann in coco_api.dataset["annotations"]:
+        cat_id = ann["category_id"]
+        if cat_id not in cat_id_2_cont_id:
+            continue
+        cont_id = cat_id_2_cont_id[cat_id]
+        ann["category_id"] = cont_id_2_cat_id[cont_id]
+        anns.append(ann)
+    coco_api.dataset["annotations"] = anns
+    # recreate index
+    coco_api.createIndex()
+
+
+def maybe_filter_and_map_categories_cocoapi(dataset_name, coco_api):
+    meta = MetadataCatalog.get(dataset_name)
+    category_id_map = meta.thing_dataset_id_to_contiguous_id
+    # map categories
+    cats = []
+    for cat in coco_api.dataset["categories"]:
+        cat_id = cat["id"]
+        if cat_id not in category_id_map:
+            continue
+        cat["id"] = category_id_map[cat_id]
+        cats.append(cat)
+    coco_api.dataset["categories"] = cats
+    # map annotation categories
+    anns = []
+    for ann in coco_api.dataset["annotations"]:
+        cat_id = ann["category_id"]
+        if cat_id not in category_id_map:
+            continue
+        ann["category_id"] = category_id_map[cat_id]
+        anns.append(ann)
+    coco_api.dataset["annotations"] = anns
+    # recreate index
+    coco_api.createIndex()
+
+
+def create_video_frame_mapping(dataset_name, dataset_dicts):
+    mapping = defaultdict(dict)
+    for d in dataset_dicts:
+        video_id = d.get("video_id")
+        if video_id is None:
+            continue
+        mapping[video_id].update({d["frame_id"]: d["file_name"]})
+    MetadataCatalog.get(dataset_name).set(video_frame_mapping=mapping)
+
+
+def load_coco_json(annotations_json_file: str, image_root: str, dataset_name: str):
+    """
+    Loads a JSON file with annotations in COCO instances format.
+    Replaces `detectron2.data.datasets.coco.load_coco_json` to handle metadata
+    in a more flexible way. Postpones category mapping to a later stage to be
+    able to combine several datasets with different (but coherent) sets of
+    categories.
+
+    Args:
+
+    annotations_json_file: str
+        Path to the JSON file with annotations in COCO instances format.
+    image_root: str
+        directory that contains all the images
+    dataset_name: str
+        the name that identifies a dataset, e.g. "densepose_coco_2014_train"
+    extra_annotation_keys: Optional[List[str]]
+        If provided, these keys are used to extract additional data from
+        the annotations.
+    """
+    coco_api = _load_coco_annotations(PathManager.get_local_path(annotations_json_file))
+    _add_categories_metadata(dataset_name, coco_api.loadCats(coco_api.getCatIds()))
+    # sort indices for reproducible results
+    img_ids = sorted(coco_api.imgs.keys())
+    # imgs is a list of dicts, each looks something like:
+    # {'license': 4,
+    #  'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
+    #  'file_name': 'COCO_val2014_000000001268.jpg',
+    #  'height': 427,
+    #  'width': 640,
+    #  'date_captured': '2013-11-17 05:57:24',
+    #  'id': 1268}
+    imgs = coco_api.loadImgs(img_ids)
+    logger = logging.getLogger(__name__)
+    logger.info("Loaded {} images in COCO format from {}".format(len(imgs), annotations_json_file))
+    # anns is a list[list[dict]], where each dict is an annotation
+    # record for an object. The inner list enumerates the objects in an image
+    # and the outer list enumerates over images.
+    anns = [coco_api.imgToAnns[img_id] for img_id in img_ids]
+    _verify_annotations_have_unique_ids(annotations_json_file, anns)
+    dataset_records = _combine_images_with_annotations(dataset_name, image_root, imgs, anns)
+    return dataset_records
+
+
+def register_dataset(dataset_data: CocoDatasetInfo, datasets_root: Optional[str] = None):
+    """
+    Registers provided COCO DensePose dataset
+
+    Args:
+    dataset_data: CocoDatasetInfo
+        Dataset data
+    datasets_root: Optional[str]
+        Datasets root folder (default: None)
+    """
+    annotations_fpath = maybe_prepend_base_path(datasets_root, dataset_data.annotations_fpath)
+    images_root = maybe_prepend_base_path(datasets_root, dataset_data.images_root)
+
+    def load_annotations():
+        return load_coco_json(
+            annotations_json_file=annotations_fpath,
+            image_root=images_root,
+            dataset_name=dataset_data.name,
+        )
+
+    DatasetCatalog.register(dataset_data.name, load_annotations)
+    MetadataCatalog.get(dataset_data.name).set(
+        json_file=annotations_fpath,
+        image_root=images_root,
+        **get_metadata(DENSEPOSE_METADATA_URL_PREFIX)
+    )
+
+
+def register_datasets(
+    datasets_data: Iterable[CocoDatasetInfo], datasets_root: Optional[str] = None
+):
+    """
+    Registers provided COCO DensePose datasets
+
+    Args:
+    datasets_data: Iterable[CocoDatasetInfo]
+        An iterable of dataset datas
+    datasets_root: Optional[str]
+        Datasets root folder (default: None)
+    """
+    for dataset_data in datasets_data:
+        register_dataset(dataset_data, datasets_root)

3rdparty/densepose/data/datasets/dataset_type.py

@@ -0,0 +1,11 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from enum import Enum
+
+
+class DatasetType(Enum):
+    """
+    Dataset type, mostly used for datasets that contain data to bootstrap models on
+    """
+
+    VIDEO_LIST = "video_list"

3rdparty/densepose/data/datasets/lvis.py

@@ -0,0 +1,257 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import logging
+import os
+from typing import Any, Dict, Iterable, List, Optional
+from fvcore.common.timer import Timer
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+from detectron2.data.datasets.lvis import get_lvis_instances_meta
+from detectron2.structures import BoxMode
+from detectron2.utils.file_io import PathManager
+
+from ..utils import maybe_prepend_base_path
+from .coco import (
+    DENSEPOSE_ALL_POSSIBLE_KEYS,
+    DENSEPOSE_METADATA_URL_PREFIX,
+    CocoDatasetInfo,
+    get_metadata,
+)
+
+DATASETS = [
+    CocoDatasetInfo(
+        name="densepose_lvis_v1_ds1_train_v1",
+        images_root="coco_",
+        annotations_fpath="lvis/densepose_lvis_v1_ds1_train_v1.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_lvis_v1_ds1_val_v1",
+        images_root="coco_",
+        annotations_fpath="lvis/densepose_lvis_v1_ds1_val_v1.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_lvis_v1_ds2_train_v1",
+        images_root="coco_",
+        annotations_fpath="lvis/densepose_lvis_v1_ds2_train_v1.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_lvis_v1_ds2_val_v1",
+        images_root="coco_",
+        annotations_fpath="lvis/densepose_lvis_v1_ds2_val_v1.json",
+    ),
+    CocoDatasetInfo(
+        name="densepose_lvis_v1_ds1_val_animals_100",
+        images_root="coco_",
+        annotations_fpath="lvis/densepose_lvis_v1_val_animals_100_v2.json",
+    ),
+]
+
+
+def _load_lvis_annotations(json_file: str):
+    """
+    Load COCO annotations from a JSON file
+
+    Args:
+        json_file: str
+            Path to the file to load annotations from
+    Returns:
+        Instance of `pycocotools.coco.COCO` that provides access to annotations
+        data
+    """
+    from lvis import LVIS
+
+    json_file = PathManager.get_local_path(json_file)
+    logger = logging.getLogger(__name__)
+    timer = Timer()
+    lvis_api = LVIS(json_file)
+    if timer.seconds() > 1:
+        logger.info("Loading {} takes {:.2f} seconds.".format(json_file, timer.seconds()))
+    return lvis_api
+
+
+def _add_categories_metadata(dataset_name: str) -> None:
+    metadict = get_lvis_instances_meta(dataset_name)
+    categories = metadict["thing_classes"]
+    metadata = MetadataCatalog.get(dataset_name)
+    metadata.categories = {i + 1: categories[i] for i in range(len(categories))}
+    logger = logging.getLogger(__name__)
+    logger.info(f"Dataset {dataset_name} has {len(categories)} categories")
+
+
+def _verify_annotations_have_unique_ids(json_file: str, anns: List[List[Dict[str, Any]]]) -> None:
+    ann_ids = [ann["id"] for anns_per_image in anns for ann in anns_per_image]
+    assert len(set(ann_ids)) == len(ann_ids), "Annotation ids in '{}' are not unique!".format(
+        json_file
+    )
+
+
+def _maybe_add_bbox(obj: Dict[str, Any], ann_dict: Dict[str, Any]) -> None:
+    if "bbox" not in ann_dict:
+        return
+    obj["bbox"] = ann_dict["bbox"]
+    obj["bbox_mode"] = BoxMode.XYWH_ABS
+
+
+def _maybe_add_segm(obj: Dict[str, Any], ann_dict: Dict[str, Any]) -> None:
+    if "segmentation" not in ann_dict:
+        return
+    segm = ann_dict["segmentation"]
+    if not isinstance(segm, dict):
+        # filter out invalid polygons (< 3 points)
+        segm = [poly for poly in segm if len(poly) % 2 == 0 and len(poly) >= 6]
+        if len(segm) == 0:
+            return
+    obj["segmentation"] = segm
+
+
+def _maybe_add_keypoints(obj: Dict[str, Any], ann_dict: Dict[str, Any]) -> None:
+    if "keypoints" not in ann_dict:
+        return
+    keypts = ann_dict["keypoints"]  # list[int]
+    for idx, v in enumerate(keypts):
+        if idx % 3 != 2:
+            # COCO's segmentation coordinates are floating points in [0, H or W],
+            # but keypoint coordinates are integers in [0, H-1 or W-1]
+            # Therefore we assume the coordinates are "pixel indices" and
+            # add 0.5 to convert to floating point coordinates.
+            keypts[idx] = v + 0.5
+    obj["keypoints"] = keypts
+
+
+def _maybe_add_densepose(obj: Dict[str, Any], ann_dict: Dict[str, Any]) -> None:
+    for key in DENSEPOSE_ALL_POSSIBLE_KEYS:
+        if key in ann_dict:
+            obj[key] = ann_dict[key]
+
+
+def _combine_images_with_annotations(
+    dataset_name: str,
+    image_root: str,
+    img_datas: Iterable[Dict[str, Any]],
+    ann_datas: Iterable[Iterable[Dict[str, Any]]],
+):
+
+    dataset_dicts = []
+
+    def get_file_name(img_root, img_dict):
+        # Determine the path including the split folder ("train2017", "val2017", "test2017") from
+        # the coco_url field. Example:
+        #   'coco_url': 'http://images.cocodataset.org/train2017/000000155379.jpg'
+        split_folder, file_name = img_dict["coco_url"].split("/")[-2:]
+        return os.path.join(img_root + split_folder, file_name)
+
+    for img_dict, ann_dicts in zip(img_datas, ann_datas):
+        record = {}
+        record["file_name"] = get_file_name(image_root, img_dict)
+        record["height"] = img_dict["height"]
+        record["width"] = img_dict["width"]
+        record["not_exhaustive_category_ids"] = img_dict.get("not_exhaustive_category_ids", [])
+        record["neg_category_ids"] = img_dict.get("neg_category_ids", [])
+        record["image_id"] = img_dict["id"]
+        record["dataset"] = dataset_name
+
+        objs = []
+        for ann_dict in ann_dicts:
+            assert ann_dict["image_id"] == record["image_id"]
+            obj = {}
+            _maybe_add_bbox(obj, ann_dict)
+            obj["iscrowd"] = ann_dict.get("iscrowd", 0)
+            obj["category_id"] = ann_dict["category_id"]
+            _maybe_add_segm(obj, ann_dict)
+            _maybe_add_keypoints(obj, ann_dict)
+            _maybe_add_densepose(obj, ann_dict)
+            objs.append(obj)
+        record["annotations"] = objs
+        dataset_dicts.append(record)
+    return dataset_dicts
+
+
+def load_lvis_json(annotations_json_file: str, image_root: str, dataset_name: str):
+    """
+    Loads a JSON file with annotations in LVIS instances format.
+    Replaces `detectron2.data.datasets.coco.load_lvis_json` to handle metadata
+    in a more flexible way. Postpones category mapping to a later stage to be
+    able to combine several datasets with different (but coherent) sets of
+    categories.
+
+    Args:
+
+    annotations_json_file: str
+        Path to the JSON file with annotations in COCO instances format.
+    image_root: str
+        directory that contains all the images
+    dataset_name: str
+        the name that identifies a dataset, e.g. "densepose_coco_2014_train"
+    extra_annotation_keys: Optional[List[str]]
+        If provided, these keys are used to extract additional data from
+        the annotations.
+    """
+    lvis_api = _load_lvis_annotations(PathManager.get_local_path(annotations_json_file))
+
+    _add_categories_metadata(dataset_name)
+
+    # sort indices for reproducible results
+    img_ids = sorted(lvis_api.imgs.keys())
+    # imgs is a list of dicts, each looks something like:
+    # {'license': 4,
+    #  'url': 'http://farm6.staticflickr.com/5454/9413846304_881d5e5c3b_z.jpg',
+    #  'file_name': 'COCO_val2014_000000001268.jpg',
+    #  'height': 427,
+    #  'width': 640,
+    #  'date_captured': '2013-11-17 05:57:24',
+    #  'id': 1268}
+    imgs = lvis_api.load_imgs(img_ids)
+    logger = logging.getLogger(__name__)
+    logger.info("Loaded {} images in LVIS format from {}".format(len(imgs), annotations_json_file))
+    # anns is a list[list[dict]], where each dict is an annotation
+    # record for an object. The inner list enumerates the objects in an image
+    # and the outer list enumerates over images.
+    anns = [lvis_api.img_ann_map[img_id] for img_id in img_ids]
+
+    _verify_annotations_have_unique_ids(annotations_json_file, anns)
+    dataset_records = _combine_images_with_annotations(dataset_name, image_root, imgs, anns)
+    return dataset_records
+
+
+def register_dataset(dataset_data: CocoDatasetInfo, datasets_root: Optional[str] = None) -> None:
+    """
+    Registers provided LVIS DensePose dataset
+
+    Args:
+        dataset_data: CocoDatasetInfo
+            Dataset data
+        datasets_root: Optional[str]
+            Datasets root folder (default: None)
+    """
+    annotations_fpath = maybe_prepend_base_path(datasets_root, dataset_data.annotations_fpath)
+    images_root = maybe_prepend_base_path(datasets_root, dataset_data.images_root)
+
+    def load_annotations():
+        return load_lvis_json(
+            annotations_json_file=annotations_fpath,
+            image_root=images_root,
+            dataset_name=dataset_data.name,
+        )
+
+    DatasetCatalog.register(dataset_data.name, load_annotations)
+    MetadataCatalog.get(dataset_data.name).set(
+        json_file=annotations_fpath,
+        image_root=images_root,
+        evaluator_type="lvis",
+        **get_metadata(DENSEPOSE_METADATA_URL_PREFIX),
+    )
+
+
+def register_datasets(
+    datasets_data: Iterable[CocoDatasetInfo], datasets_root: Optional[str] = None
+) -> None:
+    """
+    Registers provided LVIS DensePose datasets
+
+    Args:
+        datasets_data: Iterable[CocoDatasetInfo]
+            An iterable of dataset datas
+        datasets_root: Optional[str]
+            Datasets root folder (default: None)
+    """
+    for dataset_data in datasets_data:
+        register_dataset(dataset_data, datasets_root)

3rdparty/densepose/data/image_list_dataset.py

@@ -0,0 +1,72 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import logging
+import numpy as np
+from typing import Any, Callable, Dict, List, Optional, Union
+import torch
+from torch.utils.data.dataset import Dataset
+
+from detectron2.data.detection_utils import read_image
+
+ImageTransform = Callable[[torch.Tensor], torch.Tensor]
+
+
+class ImageListDataset(Dataset):
+    """
+    Dataset that provides images from a list.
+    """
+
+    _EMPTY_IMAGE = torch.empty((0, 3, 1, 1))
+
+    def __init__(
+        self,
+        image_list: List[str],
+        category_list: Union[str, List[str], None] = None,
+        transform: Optional[ImageTransform] = None,
+    ):
+        """
+        Args:
+            image_list (List[str]): list of paths to image files
+            category_list (Union[str, List[str], None]): list of animal categories for
+                each image. If it is a string, or None, this applies to all images
+        """
+        if type(category_list) == list:
+            self.category_list = category_list
+        else:
+            self.category_list = [category_list] * len(image_list)
+        assert len(image_list) == len(
+            self.category_list
+        ), "length of image and category lists must be equal"
+        self.image_list = image_list
+        self.transform = transform
+
+    def __getitem__(self, idx: int) -> Dict[str, Any]:
+        """
+        Gets selected images from the list
+
+        Args:
+            idx (int): video index in the video list file
+        Returns:
+            A dictionary containing two keys:
+                images (torch.Tensor): tensor of size [N, 3, H, W] (N = 1, or 0 for _EMPTY_IMAGE)
+                categories (List[str]): categories of the frames
+        """
+        categories = [self.category_list[idx]]
+        fpath = self.image_list[idx]
+        transform = self.transform
+
+        try:
+            image = torch.from_numpy(np.ascontiguousarray(read_image(fpath, format="BGR")))
+            image = image.permute(2, 0, 1).unsqueeze(0).float()  # HWC -> NCHW
+            if transform is not None:
+                image = transform(image)
+            return {"images": image, "categories": categories}
+        except (OSError, RuntimeError) as e:
+            logger = logging.getLogger(__name__)
+            logger.warning(f"Error opening image file container {fpath}: {e}")
+
+        return {"images": self._EMPTY_IMAGE, "categories": []}
+
+    def __len__(self):
+        return len(self.image_list)

3rdparty/densepose/data/inference_based_loader.py

@@ -0,0 +1,172 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+from typing import Any, Callable, Dict, Iterable, Iterator, List, Optional, Tuple
+import torch
+from torch import nn
+
+SampledData = Any
+ModelOutput = Any
+
+
+def _grouper(iterable: Iterable[Any], n: int, fillvalue=None) -> Iterator[Tuple[Any]]:
+    """
+    Group elements of an iterable by chunks of size `n`, e.g.
+    grouper(range(9), 4) ->
+        (0, 1, 2, 3), (4, 5, 6, 7), (8, None, None, None)
+    """
+    it = iter(iterable)
+    while True:
+        values = []
+        for _ in range(n):
+            try:
+                value = next(it)
+            except StopIteration:
+                if values:
+                    values.extend([fillvalue] * (n - len(values)))
+                    yield tuple(values)
+                return
+            values.append(value)
+        yield tuple(values)
+
+
+class ScoreBasedFilter:
+    """
+    Filters entries in model output based on their scores
+    Discards all entries with score less than the specified minimum
+    """
+
+    def __init__(self, min_score: float = 0.8):
+        self.min_score = min_score
+
+    def __call__(self, model_output: ModelOutput) -> ModelOutput:
+        for model_output_i in model_output:
+            instances = model_output_i["instances"]
+            if not instances.has("scores"):
+                continue
+            instances_filtered = instances[instances.scores >= self.min_score]
+            model_output_i["instances"] = instances_filtered
+        return model_output
+
+
+class InferenceBasedLoader:
+    """
+    Data loader based on results inferred by a model. Consists of:
+     - a data loader that provides batches of images
+     - a model that is used to infer the results
+     - a data sampler that converts inferred results to annotations
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        data_loader: Iterable[List[Dict[str, Any]]],
+        data_sampler: Optional[Callable[[ModelOutput], List[SampledData]]] = None,
+        data_filter: Optional[Callable[[ModelOutput], ModelOutput]] = None,
+        shuffle: bool = True,
+        batch_size: int = 4,
+        inference_batch_size: int = 4,
+        drop_last: bool = False,
+        category_to_class_mapping: Optional[dict] = None,
+    ):
+        """
+        Constructor
+
+        Args:
+          model (torch.nn.Module): model used to produce data
+          data_loader (Iterable[List[Dict[str, Any]]]): iterable that provides
+            dictionaries with "images" and "categories" fields to perform inference on
+          data_sampler (Callable: ModelOutput -> SampledData): functor
+              that produces annotation data from inference results;
+              (optional, default: None)
+          data_filter (Callable: ModelOutput -> ModelOutput): filter
+              that selects model outputs for further processing
+              (optional, default: None)
+          shuffle (bool): if True, the input images get shuffled
+          batch_size (int): batch size for the produced annotation data
+          inference_batch_size (int): batch size for input images
+          drop_last (bool): if True, drop the last batch if it is undersized
+          category_to_class_mapping (dict): category to class mapping
+        """
+        self.model = model
+        self.model.eval()
+        self.data_loader = data_loader
+        self.data_sampler = data_sampler
+        self.data_filter = data_filter
+        self.shuffle = shuffle
+        self.batch_size = batch_size
+        self.inference_batch_size = inference_batch_size
+        self.drop_last = drop_last
+        if category_to_class_mapping is not None:
+            self.category_to_class_mapping = category_to_class_mapping
+        else:
+            self.category_to_class_mapping = {}
+
+    def __iter__(self) -> Iterator[List[SampledData]]:
+        for batch in self.data_loader:
+            # batch : List[Dict[str: Tensor[N, C, H, W], str: Optional[str]]]
+            # images_batch : Tensor[N, C, H, W]
+            # image : Tensor[C, H, W]
+            images_and_categories = [
+                {"image": image, "category": category}
+                for element in batch
+                for image, category in zip(element["images"], element["categories"])
+            ]
+            if not images_and_categories:
+                continue
+            if self.shuffle:
+                random.shuffle(images_and_categories)
+            yield from self._produce_data(images_and_categories)  # pyre-ignore[6]
+
+    def _produce_data(
+        self, images_and_categories: List[Tuple[torch.Tensor, Optional[str]]]
+    ) -> Iterator[List[SampledData]]:
+        """
+        Produce batches of data from images
+
+        Args:
+          images_and_categories (List[Tuple[torch.Tensor, Optional[str]]]):
+            list of images and corresponding categories to process
+
+        Returns:
+          Iterator over batches of data sampled from model outputs
+        """
+        data_batches: List[SampledData] = []
+        category_to_class_mapping = self.category_to_class_mapping
+        batched_images_and_categories = _grouper(images_and_categories, self.inference_batch_size)
+        for batch in batched_images_and_categories:
+            batch = [
+                {
+                    "image": image_and_category["image"].to(self.model.device),
+                    "category": image_and_category["category"],
+                }
+                for image_and_category in batch
+                if image_and_category is not None
+            ]
+            if not batch:
+                continue
+            with torch.no_grad():
+                model_output = self.model(batch)
+            for model_output_i, batch_i in zip(model_output, batch):
+                assert len(batch_i["image"].shape) == 3
+                model_output_i["image"] = batch_i["image"]
+                instance_class = category_to_class_mapping.get(batch_i["category"], 0)
+                model_output_i["instances"].dataset_classes = torch.tensor(
+                    [instance_class] * len(model_output_i["instances"])
+                )
+            model_output_filtered = (
+                model_output if self.data_filter is None else self.data_filter(model_output)
+            )
+            data = (
+                model_output_filtered
+                if self.data_sampler is None
+                else self.data_sampler(model_output_filtered)
+            )
+            for data_i in data:
+                if len(data_i["instances"]):
+                    data_batches.append(data_i)
+            if len(data_batches) >= self.batch_size:
+                yield data_batches[: self.batch_size]
+                data_batches = data_batches[self.batch_size :]
+        if not self.drop_last and data_batches:
+            yield data_batches

3rdparty/densepose/data/meshes/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+from . import builtin
+
+__all__ = [k for k in globals().keys() if "builtin" not in k and not k.startswith("_")]

3rdparty/densepose/data/meshes/builtin.py

@@ -0,0 +1,101 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+from .catalog import MeshInfo, register_meshes
+
+DENSEPOSE_MESHES_DIR = "https://dl.fbaipublicfiles.com/densepose/meshes/"
+
+MESHES = [
+    MeshInfo(
+        name="smpl_27554",
+        data="smpl_27554.pkl",
+        geodists="geodists/geodists_smpl_27554.pkl",
+        symmetry="symmetry/symmetry_smpl_27554.pkl",
+        texcoords="texcoords/texcoords_smpl_27554.pkl",
+    ),
+    MeshInfo(
+        name="chimp_5029",
+        data="chimp_5029.pkl",
+        geodists="geodists/geodists_chimp_5029.pkl",
+        symmetry="symmetry/symmetry_chimp_5029.pkl",
+        texcoords="texcoords/texcoords_chimp_5029.pkl",
+    ),
+    MeshInfo(
+        name="cat_5001",
+        data="cat_5001.pkl",
+        geodists="geodists/geodists_cat_5001.pkl",
+        symmetry="symmetry/symmetry_cat_5001.pkl",
+        texcoords="texcoords/texcoords_cat_5001.pkl",
+    ),
+    MeshInfo(
+        name="cat_7466",
+        data="cat_7466.pkl",
+        geodists="geodists/geodists_cat_7466.pkl",
+        symmetry="symmetry/symmetry_cat_7466.pkl",
+        texcoords="texcoords/texcoords_cat_7466.pkl",
+    ),
+    MeshInfo(
+        name="sheep_5004",
+        data="sheep_5004.pkl",
+        geodists="geodists/geodists_sheep_5004.pkl",
+        symmetry="symmetry/symmetry_sheep_5004.pkl",
+        texcoords="texcoords/texcoords_sheep_5004.pkl",
+    ),
+    MeshInfo(
+        name="zebra_5002",
+        data="zebra_5002.pkl",
+        geodists="geodists/geodists_zebra_5002.pkl",
+        symmetry="symmetry/symmetry_zebra_5002.pkl",
+        texcoords="texcoords/texcoords_zebra_5002.pkl",
+    ),
+    MeshInfo(
+        name="horse_5004",
+        data="horse_5004.pkl",
+        geodists="geodists/geodists_horse_5004.pkl",
+        symmetry="symmetry/symmetry_horse_5004.pkl",
+        texcoords="texcoords/texcoords_zebra_5002.pkl",
+    ),
+    MeshInfo(
+        name="giraffe_5002",
+        data="giraffe_5002.pkl",
+        geodists="geodists/geodists_giraffe_5002.pkl",
+        symmetry="symmetry/symmetry_giraffe_5002.pkl",
+        texcoords="texcoords/texcoords_giraffe_5002.pkl",
+    ),
+    MeshInfo(
+        name="elephant_5002",
+        data="elephant_5002.pkl",
+        geodists="geodists/geodists_elephant_5002.pkl",
+        symmetry="symmetry/symmetry_elephant_5002.pkl",
+        texcoords="texcoords/texcoords_elephant_5002.pkl",
+    ),
+    MeshInfo(
+        name="dog_5002",
+        data="dog_5002.pkl",
+        geodists="geodists/geodists_dog_5002.pkl",
+        symmetry="symmetry/symmetry_dog_5002.pkl",
+        texcoords="texcoords/texcoords_dog_5002.pkl",
+    ),
+    MeshInfo(
+        name="dog_7466",
+        data="dog_7466.pkl",
+        geodists="geodists/geodists_dog_7466.pkl",
+        symmetry="symmetry/symmetry_dog_7466.pkl",
+        texcoords="texcoords/texcoords_dog_7466.pkl",
+    ),
+    MeshInfo(
+        name="cow_5002",
+        data="cow_5002.pkl",
+        geodists="geodists/geodists_cow_5002.pkl",
+        symmetry="symmetry/symmetry_cow_5002.pkl",
+        texcoords="texcoords/texcoords_cow_5002.pkl",
+    ),
+    MeshInfo(
+        name="bear_4936",
+        data="bear_4936.pkl",
+        geodists="geodists/geodists_bear_4936.pkl",
+        symmetry="symmetry/symmetry_bear_4936.pkl",
+        texcoords="texcoords/texcoords_bear_4936.pkl",
+    ),
+]
+
+register_meshes(MESHES, DENSEPOSE_MESHES_DIR)

3rdparty/densepose/data/meshes/catalog.py

@@ -0,0 +1,71 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import logging
+from collections import UserDict
+from dataclasses import dataclass
+from typing import Iterable, Optional
+
+from ..utils import maybe_prepend_base_path
+
+
+@dataclass
+class MeshInfo:
+    name: str
+    data: str
+    geodists: Optional[str] = None
+    symmetry: Optional[str] = None
+    texcoords: Optional[str] = None
+
+
+class _MeshCatalog(UserDict):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.mesh_ids = {}
+        self.mesh_names = {}
+        self.max_mesh_id = -1
+
+    def __setitem__(self, key, value):
+        if key in self:
+            logger = logging.getLogger(__name__)
+            logger.warning(
+                f"Overwriting mesh catalog entry '{key}': old value {self[key]}"
+                f", new value {value}"
+            )
+            mesh_id = self.mesh_ids[key]
+        else:
+            self.max_mesh_id += 1
+            mesh_id = self.max_mesh_id
+        super().__setitem__(key, value)
+        self.mesh_ids[key] = mesh_id
+        self.mesh_names[mesh_id] = key
+
+    def get_mesh_id(self, shape_name: str) -> int:
+        return self.mesh_ids[shape_name]
+
+    def get_mesh_name(self, mesh_id: int) -> str:
+        return self.mesh_names[mesh_id]
+
+
+MeshCatalog = _MeshCatalog()
+
+
+def register_mesh(mesh_info: MeshInfo, base_path: Optional[str]) -> None:
+    geodists, symmetry, texcoords = mesh_info.geodists, mesh_info.symmetry, mesh_info.texcoords
+    if geodists:
+        geodists = maybe_prepend_base_path(base_path, geodists)
+    if symmetry:
+        symmetry = maybe_prepend_base_path(base_path, symmetry)
+    if texcoords:
+        texcoords = maybe_prepend_base_path(base_path, texcoords)
+    MeshCatalog[mesh_info.name] = MeshInfo(
+        name=mesh_info.name,
+        data=maybe_prepend_base_path(base_path, mesh_info.data),
+        geodists=geodists,
+        symmetry=symmetry,
+        texcoords=texcoords,
+    )
+
+
+def register_meshes(mesh_infos: Iterable[MeshInfo], base_path: Optional[str]) -> None:
+    for mesh_info in mesh_infos:
+        register_mesh(mesh_info, base_path)

3rdparty/densepose/data/samplers/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .densepose_uniform import DensePoseUniformSampler
+from .densepose_confidence_based import DensePoseConfidenceBasedSampler
+from .densepose_cse_uniform import DensePoseCSEUniformSampler
+from .densepose_cse_confidence_based import DensePoseCSEConfidenceBasedSampler
+from .mask_from_densepose import MaskFromDensePoseSampler
+from .prediction_to_gt import PredictionToGroundTruthSampler

3rdparty/densepose/data/samplers/densepose_base.py

@@ -0,0 +1,203 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any, Dict, List, Tuple
+import torch
+from torch.nn import functional as F
+
+from detectron2.structures import BoxMode, Instances
+
+from densepose.converters import ToChartResultConverter
+from densepose.converters.base import IntTupleBox, make_int_box
+from densepose.structures import DensePoseDataRelative, DensePoseList
+
+
+class DensePoseBaseSampler:
+    """
+    Base DensePose sampler to produce DensePose data from DensePose predictions.
+    Samples for each class are drawn according to some distribution over all pixels estimated
+    to belong to that class.
+    """
+
+    def __init__(self, count_per_class: int = 8):
+        """
+        Constructor
+
+        Args:
+          count_per_class (int): the sampler produces at most `count_per_class`
+              samples for each category
+        """
+        self.count_per_class = count_per_class
+
+    def __call__(self, instances: Instances) -> DensePoseList:
+        """
+        Convert DensePose predictions (an instance of `DensePoseChartPredictorOutput`)
+        into DensePose annotations data (an instance of `DensePoseList`)
+        """
+        boxes_xyxy_abs = instances.pred_boxes.tensor.clone().cpu()
+        boxes_xywh_abs = BoxMode.convert(boxes_xyxy_abs, BoxMode.XYXY_ABS, BoxMode.XYWH_ABS)
+        dp_datas = []
+        for i in range(len(boxes_xywh_abs)):
+            annotation_i = self._sample(instances[i], make_int_box(boxes_xywh_abs[i]))
+            annotation_i[DensePoseDataRelative.S_KEY] = self._resample_mask(  # pyre-ignore[6]
+                instances[i].pred_densepose
+            )
+            dp_datas.append(DensePoseDataRelative(annotation_i))
+        # create densepose annotations on CPU
+        dp_list = DensePoseList(dp_datas, boxes_xyxy_abs, instances.image_size)
+        return dp_list
+
+    def _sample(self, instance: Instances, bbox_xywh: IntTupleBox) -> Dict[str, List[Any]]:
+        """
+        Sample DensPoseDataRelative from estimation results
+        """
+        labels, dp_result = self._produce_labels_and_results(instance)
+        annotation = {
+            DensePoseDataRelative.X_KEY: [],
+            DensePoseDataRelative.Y_KEY: [],
+            DensePoseDataRelative.U_KEY: [],
+            DensePoseDataRelative.V_KEY: [],
+            DensePoseDataRelative.I_KEY: [],
+        }
+        n, h, w = dp_result.shape
+        for part_id in range(1, DensePoseDataRelative.N_PART_LABELS + 1):
+            # indices - tuple of 3 1D tensors of size k
+            # 0: index along the first dimension N
+            # 1: index along H dimension
+            # 2: index along W dimension
+            indices = torch.nonzero(labels.expand(n, h, w) == part_id, as_tuple=True)
+            # values - an array of size [n, k]
+            # n: number of channels (U, V, confidences)
+            # k: number of points labeled with part_id
+            values = dp_result[indices].view(n, -1)
+            k = values.shape[1]
+            count = min(self.count_per_class, k)
+            if count <= 0:
+                continue
+            index_sample = self._produce_index_sample(values, count)
+            sampled_values = values[:, index_sample]
+            sampled_y = indices[1][index_sample] + 0.5
+            sampled_x = indices[2][index_sample] + 0.5
+            # prepare / normalize data
+            x = (sampled_x / w * 256.0).cpu().tolist()
+            y = (sampled_y / h * 256.0).cpu().tolist()
+            u = sampled_values[0].clamp(0, 1).cpu().tolist()
+            v = sampled_values[1].clamp(0, 1).cpu().tolist()
+            fine_segm_labels = [part_id] * count
+            # extend annotations
+            annotation[DensePoseDataRelative.X_KEY].extend(x)
+            annotation[DensePoseDataRelative.Y_KEY].extend(y)
+            annotation[DensePoseDataRelative.U_KEY].extend(u)
+            annotation[DensePoseDataRelative.V_KEY].extend(v)
+            annotation[DensePoseDataRelative.I_KEY].extend(fine_segm_labels)
+        return annotation
+
+    def _produce_index_sample(self, values: torch.Tensor, count: int):
+        """
+        Abstract method to produce a sample of indices to select data
+        To be implemented in descendants
+
+        Args:
+            values (torch.Tensor): an array of size [n, k] that contains
+                estimated values (U, V, confidences);
+                n: number of channels (U, V, confidences)
+                k: number of points labeled with part_id
+            count (int): number of samples to produce, should be positive and <= k
+
+        Return:
+            list(int): indices of values (along axis 1) selected as a sample
+        """
+        raise NotImplementedError
+
+    def _produce_labels_and_results(self, instance: Instances) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Method to get labels and DensePose results from an instance
+
+        Args:
+            instance (Instances): an instance of `DensePoseChartPredictorOutput`
+
+        Return:
+            labels (torch.Tensor): shape [H, W], DensePose segmentation labels
+            dp_result (torch.Tensor): shape [2, H, W], stacked DensePose results u and v
+        """
+        converter = ToChartResultConverter
+        chart_result = converter.convert(instance.pred_densepose, instance.pred_boxes)
+        labels, dp_result = chart_result.labels.cpu(), chart_result.uv.cpu()
+        return labels, dp_result
+
+    def _resample_mask(self, output: Any) -> torch.Tensor:
+        """
+        Convert DensePose predictor output to segmentation annotation - tensors of size
+        (256, 256) and type `int64`.
+
+        Args:
+            output: DensePose predictor output with the following attributes:
+             - coarse_segm: tensor of size [N, D, H, W] with unnormalized coarse
+               segmentation scores
+             - fine_segm: tensor of size [N, C, H, W] with unnormalized fine
+               segmentation scores
+        Return:
+            Tensor of size (S, S) and type `int64` with coarse segmentation annotations,
+            where S = DensePoseDataRelative.MASK_SIZE
+        """
+        sz = DensePoseDataRelative.MASK_SIZE
+        S = (
+            F.interpolate(output.coarse_segm, (sz, sz), mode="bilinear", align_corners=False)
+            .argmax(dim=1)
+            .long()
+        )
+        I = (
+            (
+                F.interpolate(
+                    output.fine_segm,
+                    (sz, sz),
+                    mode="bilinear",
+                    align_corners=False,
+                ).argmax(dim=1)
+                * (S > 0).long()
+            )
+            .squeeze()
+            .cpu()
+        )
+        # Map fine segmentation results to coarse segmentation ground truth
+        # TODO: extract this into separate classes
+        # coarse segmentation: 1 = Torso, 2 = Right Hand, 3 = Left Hand,
+        # 4 = Left Foot, 5 = Right Foot, 6 = Upper Leg Right, 7 = Upper Leg Left,
+        # 8 = Lower Leg Right, 9 = Lower Leg Left, 10 = Upper Arm Left,
+        # 11 = Upper Arm Right, 12 = Lower Arm Left, 13 = Lower Arm Right,
+        # 14 = Head
+        # fine segmentation: 1, 2 = Torso, 3 = Right Hand, 4 = Left Hand,
+        # 5 = Left Foot, 6 = Right Foot, 7, 9 = Upper Leg Right,
+        # 8, 10 = Upper Leg Left, 11, 13 = Lower Leg Right,
+        # 12, 14 = Lower Leg Left, 15, 17 = Upper Arm Left,
+        # 16, 18 = Upper Arm Right, 19, 21 = Lower Arm Left,
+        # 20, 22 = Lower Arm Right, 23, 24 = Head
+        FINE_TO_COARSE_SEGMENTATION = {
+            1: 1,
+            2: 1,
+            3: 2,
+            4: 3,
+            5: 4,
+            6: 5,
+            7: 6,
+            8: 7,
+            9: 6,
+            10: 7,
+            11: 8,
+            12: 9,
+            13: 8,
+            14: 9,
+            15: 10,
+            16: 11,
+            17: 10,
+            18: 11,
+            19: 12,
+            20: 13,
+            21: 12,
+            22: 13,
+            23: 14,
+            24: 14,
+        }
+        mask = torch.zeros((sz, sz), dtype=torch.int64, device=torch.device("cpu"))
+        for i in range(DensePoseDataRelative.N_PART_LABELS):
+            mask[I == i + 1] = FINE_TO_COARSE_SEGMENTATION[i + 1]
+        return mask

3rdparty/densepose/data/samplers/densepose_confidence_based.py

@@ -0,0 +1,108 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+from typing import Optional, Tuple
+import torch
+
+from densepose.converters import ToChartResultConverterWithConfidences
+
+from .densepose_base import DensePoseBaseSampler
+
+
+class DensePoseConfidenceBasedSampler(DensePoseBaseSampler):
+    """
+    Samples DensePose data from DensePose predictions.
+    Samples for each class are drawn using confidence value estimates.
+    """
+
+    def __init__(
+        self,
+        confidence_channel: str,
+        count_per_class: int = 8,
+        search_count_multiplier: Optional[float] = None,
+        search_proportion: Optional[float] = None,
+    ):
+        """
+        Constructor
+
+        Args:
+          confidence_channel (str): confidence channel to use for sampling;
+            possible values:
+              "sigma_2": confidences for UV values
+              "fine_segm_confidence": confidences for fine segmentation
+              "coarse_segm_confidence": confidences for coarse segmentation
+            (default: "sigma_2")
+          count_per_class (int): the sampler produces at most `count_per_class`
+              samples for each category (default: 8)
+          search_count_multiplier (float or None): if not None, the total number
+              of the most confident estimates of a given class to consider is
+              defined as `min(search_count_multiplier * count_per_class, N)`,
+              where `N` is the total number of estimates of the class; cannot be
+              specified together with `search_proportion` (default: None)
+          search_proportion (float or None): if not None, the total number of the
+              of the most confident estimates of a given class to consider is
+              defined as `min(max(search_proportion * N, count_per_class), N)`,
+              where `N` is the total number of estimates of the class; cannot be
+              specified together with `search_count_multiplier` (default: None)
+        """
+        super().__init__(count_per_class)
+        self.confidence_channel = confidence_channel
+        self.search_count_multiplier = search_count_multiplier
+        self.search_proportion = search_proportion
+        assert (search_count_multiplier is None) or (search_proportion is None), (
+            f"Cannot specify both search_count_multiplier (={search_count_multiplier})"
+            f"and search_proportion (={search_proportion})"
+        )
+
+    def _produce_index_sample(self, values: torch.Tensor, count: int):
+        """
+        Produce a sample of indices to select data based on confidences
+
+        Args:
+            values (torch.Tensor): an array of size [n, k] that contains
+                estimated values (U, V, confidences);
+                n: number of channels (U, V, confidences)
+                k: number of points labeled with part_id
+            count (int): number of samples to produce, should be positive and <= k
+
+        Return:
+            list(int): indices of values (along axis 1) selected as a sample
+        """
+        k = values.shape[1]
+        if k == count:
+            index_sample = list(range(k))
+        else:
+            # take the best count * search_count_multiplier pixels,
+            # sample from them uniformly
+            # (here best = smallest variance)
+            _, sorted_confidence_indices = torch.sort(values[2])
+            if self.search_count_multiplier is not None:
+                search_count = min(int(count * self.search_count_multiplier), k)
+            elif self.search_proportion is not None:
+                search_count = min(max(int(k * self.search_proportion), count), k)
+            else:
+                search_count = min(count, k)
+            sample_from_top = random.sample(range(search_count), count)
+            index_sample = sorted_confidence_indices[:search_count][sample_from_top]
+        return index_sample
+
+    def _produce_labels_and_results(self, instance) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Method to get labels and DensePose results from an instance, with confidences
+
+        Args:
+            instance (Instances): an instance of `DensePoseChartPredictorOutputWithConfidences`
+
+        Return:
+            labels (torch.Tensor): shape [H, W], DensePose segmentation labels
+            dp_result (torch.Tensor): shape [3, H, W], DensePose results u and v
+                stacked with the confidence channel
+        """
+        converter = ToChartResultConverterWithConfidences
+        chart_result = converter.convert(instance.pred_densepose, instance.pred_boxes)
+        labels, dp_result = chart_result.labels.cpu(), chart_result.uv.cpu()
+        dp_result = torch.cat(
+            (dp_result, getattr(chart_result, self.confidence_channel)[None].cpu())
+        )
+
+        return labels, dp_result

3rdparty/densepose/data/samplers/densepose_cse_base.py

@@ -0,0 +1,139 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from typing import Any, Dict, List, Tuple
+import torch
+from torch.nn import functional as F
+
+from detectron2.config import CfgNode
+from detectron2.structures import Instances
+
+from densepose.converters.base import IntTupleBox
+from densepose.data.utils import get_class_to_mesh_name_mapping
+from densepose.modeling.cse.utils import squared_euclidean_distance_matrix
+from densepose.structures import DensePoseDataRelative
+
+from .densepose_base import DensePoseBaseSampler
+
+
+class DensePoseCSEBaseSampler(DensePoseBaseSampler):
+    """
+    Base DensePose sampler to produce DensePose data from DensePose predictions.
+    Samples for each class are drawn according to some distribution over all pixels estimated
+    to belong to that class.
+    """
+
+    def __init__(
+        self,
+        cfg: CfgNode,
+        use_gt_categories: bool,
+        embedder: torch.nn.Module,
+        count_per_class: int = 8,
+    ):
+        """
+        Constructor
+
+        Args:
+          cfg (CfgNode): the config of the model
+          embedder (torch.nn.Module): necessary to compute mesh vertex embeddings
+          count_per_class (int): the sampler produces at most `count_per_class`
+              samples for each category
+        """
+        super().__init__(count_per_class)
+        self.embedder = embedder
+        self.class_to_mesh_name = get_class_to_mesh_name_mapping(cfg)
+        self.use_gt_categories = use_gt_categories
+
+    def _sample(self, instance: Instances, bbox_xywh: IntTupleBox) -> Dict[str, List[Any]]:
+        """
+        Sample DensPoseDataRelative from estimation results
+        """
+        if self.use_gt_categories:
+            instance_class = instance.dataset_classes.tolist()[0]
+        else:
+            instance_class = instance.pred_classes.tolist()[0]
+        mesh_name = self.class_to_mesh_name[instance_class]
+
+        annotation = {
+            DensePoseDataRelative.X_KEY: [],
+            DensePoseDataRelative.Y_KEY: [],
+            DensePoseDataRelative.VERTEX_IDS_KEY: [],
+            DensePoseDataRelative.MESH_NAME_KEY: mesh_name,
+        }
+
+        mask, embeddings, other_values = self._produce_mask_and_results(instance, bbox_xywh)
+        indices = torch.nonzero(mask, as_tuple=True)
+        selected_embeddings = embeddings.permute(1, 2, 0)[indices].cpu()
+        values = other_values[:, indices[0], indices[1]]
+        k = values.shape[1]
+
+        count = min(self.count_per_class, k)
+        if count <= 0:
+            return annotation
+
+        index_sample = self._produce_index_sample(values, count)
+        closest_vertices = squared_euclidean_distance_matrix(
+            selected_embeddings[index_sample], self.embedder(mesh_name)
+        )
+        closest_vertices = torch.argmin(closest_vertices, dim=1)
+
+        sampled_y = indices[0][index_sample] + 0.5
+        sampled_x = indices[1][index_sample] + 0.5
+        # prepare / normalize data
+        _, _, w, h = bbox_xywh
+        x = (sampled_x / w * 256.0).cpu().tolist()
+        y = (sampled_y / h * 256.0).cpu().tolist()
+        # extend annotations
+        annotation[DensePoseDataRelative.X_KEY].extend(x)
+        annotation[DensePoseDataRelative.Y_KEY].extend(y)
+        annotation[DensePoseDataRelative.VERTEX_IDS_KEY].extend(closest_vertices.cpu().tolist())
+        return annotation
+
+    def _produce_mask_and_results(
+        self, instance: Instances, bbox_xywh: IntTupleBox
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Method to get labels and DensePose results from an instance
+
+        Args:
+            instance (Instances): an instance of `DensePoseEmbeddingPredictorOutput`
+            bbox_xywh (IntTupleBox): the corresponding bounding box
+
+        Return:
+            mask (torch.Tensor): shape [H, W], DensePose segmentation mask
+            embeddings (Tuple[torch.Tensor]): a tensor of shape [D, H, W],
+                DensePose CSE Embeddings
+            other_values (Tuple[torch.Tensor]): a tensor of shape [0, H, W],
+                for potential other values
+        """
+        densepose_output = instance.pred_densepose
+        S = densepose_output.coarse_segm
+        E = densepose_output.embedding
+        _, _, w, h = bbox_xywh
+        embeddings = F.interpolate(E, size=(h, w), mode="bilinear")[0]
+        coarse_segm_resized = F.interpolate(S, size=(h, w), mode="bilinear")[0]
+        mask = coarse_segm_resized.argmax(0) > 0
+        other_values = torch.empty((0, h, w), device=E.device)
+        return mask, embeddings, other_values
+
+    def _resample_mask(self, output: Any) -> torch.Tensor:
+        """
+        Convert DensePose predictor output to segmentation annotation - tensors of size
+        (256, 256) and type `int64`.
+
+        Args:
+            output: DensePose predictor output with the following attributes:
+             - coarse_segm: tensor of size [N, D, H, W] with unnormalized coarse
+               segmentation scores
+        Return:
+            Tensor of size (S, S) and type `int64` with coarse segmentation annotations,
+            where S = DensePoseDataRelative.MASK_SIZE
+        """
+        sz = DensePoseDataRelative.MASK_SIZE
+        mask = (
+            F.interpolate(output.coarse_segm, (sz, sz), mode="bilinear", align_corners=False)
+            .argmax(dim=1)
+            .long()
+            .squeeze()
+            .cpu()
+        )
+        return mask

3rdparty/densepose/data/samplers/densepose_cse_confidence_based.py

@@ -0,0 +1,119 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+from typing import Optional, Tuple
+import torch
+from torch.nn import functional as F
+
+from detectron2.config import CfgNode
+from detectron2.structures import Instances
+
+from densepose.converters.base import IntTupleBox
+
+from .densepose_cse_base import DensePoseCSEBaseSampler
+
+
+class DensePoseCSEConfidenceBasedSampler(DensePoseCSEBaseSampler):
+    """
+    Samples DensePose data from DensePose predictions.
+    Samples for each class are drawn using confidence value estimates.
+    """
+
+    def __init__(
+        self,
+        cfg: CfgNode,
+        use_gt_categories: bool,
+        embedder: torch.nn.Module,
+        confidence_channel: str,
+        count_per_class: int = 8,
+        search_count_multiplier: Optional[float] = None,
+        search_proportion: Optional[float] = None,
+    ):
+        """
+        Constructor
+
+        Args:
+          cfg (CfgNode): the config of the model
+          embedder (torch.nn.Module): necessary to compute mesh vertex embeddings
+          confidence_channel (str): confidence channel to use for sampling;
+            possible values:
+              "coarse_segm_confidence": confidences for coarse segmentation
+            (default: "coarse_segm_confidence")
+          count_per_class (int): the sampler produces at most `count_per_class`
+              samples for each category (default: 8)
+          search_count_multiplier (float or None): if not None, the total number
+              of the most confident estimates of a given class to consider is
+              defined as `min(search_count_multiplier * count_per_class, N)`,
+              where `N` is the total number of estimates of the class; cannot be
+              specified together with `search_proportion` (default: None)
+          search_proportion (float or None): if not None, the total number of the
+              of the most confident estimates of a given class to consider is
+              defined as `min(max(search_proportion * N, count_per_class), N)`,
+              where `N` is the total number of estimates of the class; cannot be
+              specified together with `search_count_multiplier` (default: None)
+        """
+        super().__init__(cfg, use_gt_categories, embedder, count_per_class)
+        self.confidence_channel = confidence_channel
+        self.search_count_multiplier = search_count_multiplier
+        self.search_proportion = search_proportion
+        assert (search_count_multiplier is None) or (search_proportion is None), (
+            f"Cannot specify both search_count_multiplier (={search_count_multiplier})"
+            f"and search_proportion (={search_proportion})"
+        )
+
+    def _produce_index_sample(self, values: torch.Tensor, count: int):
+        """
+        Produce a sample of indices to select data based on confidences
+
+        Args:
+            values (torch.Tensor): a tensor of length k that contains confidences
+                k: number of points labeled with part_id
+            count (int): number of samples to produce, should be positive and <= k
+
+        Return:
+            list(int): indices of values (along axis 1) selected as a sample
+        """
+        k = values.shape[1]
+        if k == count:
+            index_sample = list(range(k))
+        else:
+            # take the best count * search_count_multiplier pixels,
+            # sample from them uniformly
+            # (here best = smallest variance)
+            _, sorted_confidence_indices = torch.sort(values[0])
+            if self.search_count_multiplier is not None:
+                search_count = min(int(count * self.search_count_multiplier), k)
+            elif self.search_proportion is not None:
+                search_count = min(max(int(k * self.search_proportion), count), k)
+            else:
+                search_count = min(count, k)
+            sample_from_top = random.sample(range(search_count), count)
+            index_sample = sorted_confidence_indices[-search_count:][sample_from_top]
+        return index_sample
+
+    def _produce_mask_and_results(
+        self, instance: Instances, bbox_xywh: IntTupleBox
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Method to get labels and DensePose results from an instance
+
+        Args:
+            instance (Instances): an instance of
+                `DensePoseEmbeddingPredictorOutputWithConfidences`
+            bbox_xywh (IntTupleBox): the corresponding bounding box
+
+        Return:
+            mask (torch.Tensor): shape [H, W], DensePose segmentation mask
+            embeddings (Tuple[torch.Tensor]): a tensor of shape [D, H, W]
+                DensePose CSE Embeddings
+            other_values: a tensor of shape [1, H, W], DensePose CSE confidence
+        """
+        _, _, w, h = bbox_xywh
+        densepose_output = instance.pred_densepose
+        mask, embeddings, _ = super()._produce_mask_and_results(instance, bbox_xywh)
+        other_values = F.interpolate(
+            getattr(densepose_output, self.confidence_channel),
+            size=(h, w),
+            mode="bilinear",
+        )[0].cpu()
+        return mask, embeddings, other_values

3rdparty/densepose/data/samplers/densepose_cse_uniform.py

@@ -0,0 +1,12 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .densepose_cse_base import DensePoseCSEBaseSampler
+from .densepose_uniform import DensePoseUniformSampler
+
+
+class DensePoseCSEUniformSampler(DensePoseCSEBaseSampler, DensePoseUniformSampler):
+    """
+    Uniform Sampler for CSE
+    """
+
+    pass

3rdparty/densepose/data/samplers/densepose_uniform.py

@@ -0,0 +1,41 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+import torch
+
+from .densepose_base import DensePoseBaseSampler
+
+
+class DensePoseUniformSampler(DensePoseBaseSampler):
+    """
+    Samples DensePose data from DensePose predictions.
+    Samples for each class are drawn uniformly over all pixels estimated
+    to belong to that class.
+    """
+
+    def __init__(self, count_per_class: int = 8):
+        """
+        Constructor
+
+        Args:
+          count_per_class (int): the sampler produces at most `count_per_class`
+              samples for each category
+        """
+        super().__init__(count_per_class)
+
+    def _produce_index_sample(self, values: torch.Tensor, count: int):
+        """
+        Produce a uniform sample of indices to select data
+
+        Args:
+            values (torch.Tensor): an array of size [n, k] that contains
+                estimated values (U, V, confidences);
+                n: number of channels (U, V, confidences)
+                k: number of points labeled with part_id
+            count (int): number of samples to produce, should be positive and <= k
+
+        Return:
+            list(int): indices of values (along axis 1) selected as a sample
+        """
+        k = values.shape[1]
+        return random.sample(range(k), count)

3rdparty/densepose/data/samplers/mask_from_densepose.py

@@ -0,0 +1,28 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from detectron2.structures import BitMasks, Instances
+
+from densepose.converters import ToMaskConverter
+
+
+class MaskFromDensePoseSampler:
+    """
+    Produce mask GT from DensePose predictions
+    This sampler simply converts DensePose predictions to BitMasks
+    that a contain a bool tensor of the size of the input image
+    """
+
+    def __call__(self, instances: Instances) -> BitMasks:
+        """
+        Converts predicted data from `instances` into the GT mask data
+
+        Args:
+            instances (Instances): predicted results, expected to have `pred_densepose` field
+
+        Returns:
+            Boolean Tensor of the size of the input image that has non-zero
+            values at pixels that are estimated to belong to the detected object
+        """
+        return ToMaskConverter.convert(
+            instances.pred_densepose, instances.pred_boxes, instances.image_size
+        )

3rdparty/densepose/data/samplers/prediction_to_gt.py

@@ -0,0 +1,98 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional
+
+from detectron2.structures import Instances
+
+ModelOutput = Dict[str, Any]
+SampledData = Dict[str, Any]
+
+
+@dataclass
+class _Sampler:
+    """
+    Sampler registry entry that contains:
+     - src (str): source field to sample from (deleted after sampling)
+     - dst (Optional[str]): destination field to sample to, if not None
+     - func (Optional[Callable: Any -> Any]): function that performs sampling,
+         if None, reference copy is performed
+    """
+
+    src: str
+    dst: Optional[str]
+    func: Optional[Callable[[Any], Any]]
+
+
+class PredictionToGroundTruthSampler:
+    """
+    Sampler implementation that converts predictions to GT using registered
+    samplers for different fields of `Instances`.
+    """
+
+    def __init__(self, dataset_name: str = ""):
+        self.dataset_name = dataset_name
+        self._samplers = {}
+        self.register_sampler("pred_boxes", "gt_boxes", None)
+        self.register_sampler("pred_classes", "gt_classes", None)
+        # delete scores
+        self.register_sampler("scores")
+
+    def __call__(self, model_output: List[ModelOutput]) -> List[SampledData]:
+        """
+        Transform model output into ground truth data through sampling
+
+        Args:
+          model_output (Dict[str, Any]): model output
+        Returns:
+          Dict[str, Any]: sampled data
+        """
+        for model_output_i in model_output:
+            instances: Instances = model_output_i["instances"]
+            # transform data in each field
+            for _, sampler in self._samplers.items():
+                if not instances.has(sampler.src) or sampler.dst is None:
+                    continue
+                if sampler.func is None:
+                    instances.set(sampler.dst, instances.get(sampler.src))
+                else:
+                    instances.set(sampler.dst, sampler.func(instances))
+            # delete model output data that was transformed
+            for _, sampler in self._samplers.items():
+                if sampler.src != sampler.dst and instances.has(sampler.src):
+                    instances.remove(sampler.src)
+            model_output_i["dataset"] = self.dataset_name
+        return model_output
+
+    def register_sampler(
+        self,
+        prediction_attr: str,
+        gt_attr: Optional[str] = None,
+        func: Optional[Callable[[Any], Any]] = None,
+    ):
+        """
+        Register sampler for a field
+
+        Args:
+          prediction_attr (str): field to replace with a sampled value
+          gt_attr (Optional[str]): field to store the sampled value to, if not None
+          func (Optional[Callable: Any -> Any]): sampler function
+        """
+        self._samplers[(prediction_attr, gt_attr)] = _Sampler(
+            src=prediction_attr, dst=gt_attr, func=func
+        )
+
+    def remove_sampler(
+        self,
+        prediction_attr: str,
+        gt_attr: Optional[str] = None,
+    ):
+        """
+        Remove sampler for a field
+
+        Args:
+          prediction_attr (str): field to replace with a sampled value
+          gt_attr (Optional[str]): field to store the sampled value to, if not None
+        """
+        assert (prediction_attr, gt_attr) in self._samplers
+        del self._samplers[(prediction_attr, gt_attr)]

3rdparty/densepose/data/transform/__init__.py

@@ -0,0 +1,3 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .image import ImageResizeTransform

3rdparty/densepose/data/transform/image.py

@@ -0,0 +1,39 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import torch
+
+
+class ImageResizeTransform:
+    """
+    Transform that resizes images loaded from a dataset
+    (BGR data in NCHW channel order, typically uint8) to a format ready to be
+    consumed by DensePose training (BGR float32 data in NCHW channel order)
+    """
+
+    def __init__(self, min_size: int = 800, max_size: int = 1333):
+        self.min_size = min_size
+        self.max_size = max_size
+
+    def __call__(self, images: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            images (torch.Tensor): tensor of size [N, 3, H, W] that contains
+                BGR data (typically in uint8)
+        Returns:
+            images (torch.Tensor): tensor of size [N, 3, H1, W1] where
+                H1 and W1 are chosen to respect the specified min and max sizes
+                and preserve the original aspect ratio, the data channels
+                follow BGR order and the data type is `torch.float32`
+        """
+        # resize with min size
+        images = images.float()
+        min_size = min(images.shape[-2:])
+        max_size = max(images.shape[-2:])
+        scale = min(self.min_size / min_size, self.max_size / max_size)
+        images = torch.nn.functional.interpolate(
+            images,
+            scale_factor=scale,
+            mode="bilinear",
+            align_corners=False,
+        )
+        return images

3rdparty/densepose/data/utils.py

@@ -0,0 +1,38 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import os
+from typing import Dict, Optional
+
+from detectron2.config import CfgNode
+
+
+def is_relative_local_path(path: str) -> bool:
+    path_str = os.fsdecode(path)
+    return ("://" not in path_str) and not os.path.isabs(path)
+
+
+def maybe_prepend_base_path(base_path: Optional[str], path: str):
+    """
+    Prepends the provided path with a base path prefix if:
+    1) base path is not None;
+    2) path is a local path
+    """
+    if base_path is None:
+        return path
+    if is_relative_local_path(path):
+        return os.path.join(base_path, path)
+    return path
+
+
+def get_class_to_mesh_name_mapping(cfg: CfgNode) -> Dict[int, str]:
+    return {
+        int(class_id): mesh_name
+        for class_id, mesh_name in cfg.DATASETS.CLASS_TO_MESH_NAME_MAPPING.items()
+    }
+
+
+def get_category_to_class_mapping(dataset_cfg: CfgNode) -> Dict[str, int]:
+    return {
+        category: int(class_id)
+        for category, class_id in dataset_cfg.CATEGORY_TO_CLASS_MAPPING.items()
+    }

3rdparty/densepose/data/video/__init__.py

@@ -0,0 +1,17 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .frame_selector import (
+    FrameSelectionStrategy,
+    RandomKFramesSelector,
+    FirstKFramesSelector,
+    LastKFramesSelector,
+    FrameTsList,
+    FrameSelector,
+)
+
+from .video_keyframe_dataset import (
+    VideoKeyframeDataset,
+    video_list_from_file,
+    list_keyframes,
+    read_keyframes,
+)

3rdparty/densepose/data/video/frame_selector.py

@@ -0,0 +1,87 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import random
+from collections.abc import Callable
+from enum import Enum
+from typing import Callable as TCallable
+from typing import List
+
+FrameTsList = List[int]
+FrameSelector = TCallable[[FrameTsList], FrameTsList]
+
+
+class FrameSelectionStrategy(Enum):
+    """
+    Frame selection strategy used with videos:
+     - "random_k": select k random frames
+     - "first_k": select k first frames
+     - "last_k": select k last frames
+     - "all": select all frames
+    """
+
+    # fmt: off
+    RANDOM_K = "random_k"
+    FIRST_K  = "first_k"
+    LAST_K   = "last_k"
+    ALL      = "all"
+    # fmt: on
+
+
+class RandomKFramesSelector(Callable):  # pyre-ignore[39]
+    """
+    Selector that retains at most `k` random frames
+    """
+
+    def __init__(self, k: int):
+        self.k = k
+
+    def __call__(self, frame_tss: FrameTsList) -> FrameTsList:
+        """
+        Select `k` random frames
+
+        Args:
+          frames_tss (List[int]): timestamps of input frames
+        Returns:
+          List[int]: timestamps of selected frames
+        """
+        return random.sample(frame_tss, min(self.k, len(frame_tss)))
+
+
+class FirstKFramesSelector(Callable):  # pyre-ignore[39]
+    """
+    Selector that retains at most `k` first frames
+    """
+
+    def __init__(self, k: int):
+        self.k = k
+
+    def __call__(self, frame_tss: FrameTsList) -> FrameTsList:
+        """
+        Select `k` first frames
+
+        Args:
+          frames_tss (List[int]): timestamps of input frames
+        Returns:
+          List[int]: timestamps of selected frames
+        """
+        return frame_tss[: self.k]
+
+
+class LastKFramesSelector(Callable):  # pyre-ignore[39]
+    """
+    Selector that retains at most `k` last frames from video data
+    """
+
+    def __init__(self, k: int):
+        self.k = k
+
+    def __call__(self, frame_tss: FrameTsList) -> FrameTsList:
+        """
+        Select `k` last frames
+
+        Args:
+          frames_tss (List[int]): timestamps of input frames
+        Returns:
+          List[int]: timestamps of selected frames
+        """
+        return frame_tss[-self.k :]

3rdparty/densepose/data/video/video_keyframe_dataset.py

@@ -0,0 +1,300 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import csv
+import logging
+import numpy as np
+from typing import Any, Callable, Dict, List, Optional, Union
+import av
+import torch
+from torch.utils.data.dataset import Dataset
+
+from detectron2.utils.file_io import PathManager
+
+from ..utils import maybe_prepend_base_path
+from .frame_selector import FrameSelector, FrameTsList
+
+FrameList = List[av.frame.Frame]  # pyre-ignore[16]
+FrameTransform = Callable[[torch.Tensor], torch.Tensor]
+
+
+def list_keyframes(video_fpath: str, video_stream_idx: int = 0) -> FrameTsList:
+    """
+    Traverses all keyframes of a video file. Returns a list of keyframe
+    timestamps. Timestamps are counts in timebase units.
+
+    Args:
+       video_fpath (str): Video file path
+       video_stream_idx (int): Video stream index (default: 0)
+    Returns:
+       List[int]: list of keyframe timestaps (timestamp is a count in timebase
+           units)
+    """
+    try:
+        with PathManager.open(video_fpath, "rb") as io:
+            container = av.open(io, mode="r")
+            stream = container.streams.video[video_stream_idx]
+            keyframes = []
+            pts = -1
+            # Note: even though we request forward seeks for keyframes, sometimes
+            # a keyframe in backwards direction is returned. We introduce tolerance
+            # as a max count of ignored backward seeks
+            tolerance_backward_seeks = 2
+            while True:
+                try:
+                    container.seek(pts + 1, backward=False, any_frame=False, stream=stream)
+                except av.AVError as e:
+                    # the exception occurs when the video length is exceeded,
+                    # we then return whatever data we've already collected
+                    logger = logging.getLogger(__name__)
+                    logger.debug(
+                        f"List keyframes: Error seeking video file {video_fpath}, "
+                        f"video stream {video_stream_idx}, pts {pts + 1}, AV error: {e}"
+                    )
+                    return keyframes
+                except OSError as e:
+                    logger = logging.getLogger(__name__)
+                    logger.warning(
+                        f"List keyframes: Error seeking video file {video_fpath}, "
+                        f"video stream {video_stream_idx}, pts {pts + 1}, OS error: {e}"
+                    )
+                    return []
+                packet = next(container.demux(video=video_stream_idx))
+                if packet.pts is not None and packet.pts <= pts:
+                    logger = logging.getLogger(__name__)
+                    logger.warning(
+                        f"Video file {video_fpath}, stream {video_stream_idx}: "
+                        f"bad seek for packet {pts + 1} (got packet {packet.pts}), "
+                        f"tolerance {tolerance_backward_seeks}."
+                    )
+                    tolerance_backward_seeks -= 1
+                    if tolerance_backward_seeks == 0:
+                        return []
+                    pts += 1
+                    continue
+                tolerance_backward_seeks = 2
+                pts = packet.pts
+                if pts is None:
+                    return keyframes
+                if packet.is_keyframe:
+                    keyframes.append(pts)
+            return keyframes
+    except OSError as e:
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            f"List keyframes: Error opening video file container {video_fpath}, " f"OS error: {e}"
+        )
+    except RuntimeError as e:
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            f"List keyframes: Error opening video file container {video_fpath}, "
+            f"Runtime error: {e}"
+        )
+    return []
+
+
+def read_keyframes(
+    video_fpath: str, keyframes: FrameTsList, video_stream_idx: int = 0
+) -> FrameList:  # pyre-ignore[11]
+    """
+    Reads keyframe data from a video file.
+
+    Args:
+        video_fpath (str): Video file path
+        keyframes (List[int]): List of keyframe timestamps (as counts in
+            timebase units to be used in container seek operations)
+        video_stream_idx (int): Video stream index (default: 0)
+    Returns:
+        List[Frame]: list of frames that correspond to the specified timestamps
+    """
+    try:
+        with PathManager.open(video_fpath, "rb") as io:
+            container = av.open(io)
+            stream = container.streams.video[video_stream_idx]
+            frames = []
+            for pts in keyframes:
+                try:
+                    container.seek(pts, any_frame=False, stream=stream)
+                    frame = next(container.decode(video=0))
+                    frames.append(frame)
+                except av.AVError as e:
+                    logger = logging.getLogger(__name__)
+                    logger.warning(
+                        f"Read keyframes: Error seeking video file {video_fpath}, "
+                        f"video stream {video_stream_idx}, pts {pts}, AV error: {e}"
+                    )
+                    container.close()
+                    return frames
+                except OSError as e:
+                    logger = logging.getLogger(__name__)
+                    logger.warning(
+                        f"Read keyframes: Error seeking video file {video_fpath}, "
+                        f"video stream {video_stream_idx}, pts {pts}, OS error: {e}"
+                    )
+                    container.close()
+                    return frames
+                except StopIteration:
+                    logger = logging.getLogger(__name__)
+                    logger.warning(
+                        f"Read keyframes: Error decoding frame from {video_fpath}, "
+                        f"video stream {video_stream_idx}, pts {pts}"
+                    )
+                    container.close()
+                    return frames
+
+            container.close()
+            return frames
+    except OSError as e:
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            f"Read keyframes: Error opening video file container {video_fpath}, OS error: {e}"
+        )
+    except RuntimeError as e:
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            f"Read keyframes: Error opening video file container {video_fpath}, Runtime error: {e}"
+        )
+    return []
+
+
+def video_list_from_file(video_list_fpath: str, base_path: Optional[str] = None):
+    """
+    Create a list of paths to video files from a text file.
+
+    Args:
+        video_list_fpath (str): path to a plain text file with the list of videos
+        base_path (str): base path for entries from the video list (default: None)
+    """
+    video_list = []
+    with PathManager.open(video_list_fpath, "r") as io:
+        for line in io:
+            video_list.append(maybe_prepend_base_path(base_path, str(line.strip())))
+    return video_list
+
+
+def read_keyframe_helper_data(fpath: str):
+    """
+    Read keyframe data from a file in CSV format: the header should contain
+    "video_id" and "keyframes" fields. Value specifications are:
+      video_id: int
+      keyframes: list(int)
+    Example of contents:
+      video_id,keyframes
+      2,"[1,11,21,31,41,51,61,71,81]"
+
+    Args:
+        fpath (str): File containing keyframe data
+
+    Return:
+        video_id_to_keyframes (dict: int -> list(int)): for a given video ID it
+          contains a list of keyframes for that video
+    """
+    video_id_to_keyframes = {}
+    try:
+        with PathManager.open(fpath, "r") as io:
+            csv_reader = csv.reader(io)
+            header = next(csv_reader)
+            video_id_idx = header.index("video_id")
+            keyframes_idx = header.index("keyframes")
+            for row in csv_reader:
+                video_id = int(row[video_id_idx])
+                assert (
+                    video_id not in video_id_to_keyframes
+                ), f"Duplicate keyframes entry for video {fpath}"
+                video_id_to_keyframes[video_id] = (
+                    [int(v) for v in row[keyframes_idx][1:-1].split(",")]
+                    if len(row[keyframes_idx]) > 2
+                    else []
+                )
+    except Exception as e:
+        logger = logging.getLogger(__name__)
+        logger.warning(f"Error reading keyframe helper data from {fpath}: {e}")
+    return video_id_to_keyframes
+
+
+class VideoKeyframeDataset(Dataset):
+    """
+    Dataset that provides keyframes for a set of videos.
+    """
+
+    _EMPTY_FRAMES = torch.empty((0, 3, 1, 1))
+
+    def __init__(
+        self,
+        video_list: List[str],
+        category_list: Union[str, List[str], None] = None,
+        frame_selector: Optional[FrameSelector] = None,
+        transform: Optional[FrameTransform] = None,
+        keyframe_helper_fpath: Optional[str] = None,
+    ):
+        """
+        Dataset constructor
+
+        Args:
+            video_list (List[str]): list of paths to video files
+            category_list (Union[str, List[str], None]): list of animal categories for each
+                video file. If it is a string, or None, this applies to all videos
+            frame_selector (Callable: KeyFrameList -> KeyFrameList):
+                selects keyframes to process, keyframes are given by
+                packet timestamps in timebase counts. If None, all keyframes
+                are selected (default: None)
+            transform (Callable: torch.Tensor -> torch.Tensor):
+                transforms a batch of RGB images (tensors of size [B, 3, H, W]),
+                returns a tensor of the same size. If None, no transform is
+                applied (default: None)
+
+        """
+        if type(category_list) == list:
+            self.category_list = category_list
+        else:
+            self.category_list = [category_list] * len(video_list)
+        assert len(video_list) == len(
+            self.category_list
+        ), "length of video and category lists must be equal"
+        self.video_list = video_list
+        self.frame_selector = frame_selector
+        self.transform = transform
+        self.keyframe_helper_data = (
+            read_keyframe_helper_data(keyframe_helper_fpath)
+            if keyframe_helper_fpath is not None
+            else None
+        )
+
+    def __getitem__(self, idx: int) -> Dict[str, Any]:
+        """
+        Gets selected keyframes from a given video
+
+        Args:
+            idx (int): video index in the video list file
+        Returns:
+            A dictionary containing two keys:
+                images (torch.Tensor): tensor of size [N, H, W, 3] or of size
+                    defined by the transform that contains keyframes data
+                categories (List[str]): categories of the frames
+        """
+        categories = [self.category_list[idx]]
+        fpath = self.video_list[idx]
+        keyframes = (
+            list_keyframes(fpath)
+            if self.keyframe_helper_data is None or idx not in self.keyframe_helper_data
+            else self.keyframe_helper_data[idx]
+        )
+        transform = self.transform
+        frame_selector = self.frame_selector
+        if not keyframes:
+            return {"images": self._EMPTY_FRAMES, "categories": []}
+        if frame_selector is not None:
+            keyframes = frame_selector(keyframes)
+        frames = read_keyframes(fpath, keyframes)
+        if not frames:
+            return {"images": self._EMPTY_FRAMES, "categories": []}
+        frames = np.stack([frame.to_rgb().to_ndarray() for frame in frames])
+        frames = torch.as_tensor(frames, device=torch.device("cpu"))
+        frames = frames[..., [2, 1, 0]]  # RGB -> BGR
+        frames = frames.permute(0, 3, 1, 2).float()  # NHWC -> NCHW
+        if transform is not None:
+            frames = transform(frames)
+        return {"images": frames, "categories": categories}
+
+    def __len__(self):
+        return len(self.video_list)

3rdparty/densepose/engine/__init__.py

@@ -0,0 +1,3 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .trainer import Trainer

3rdparty/densepose/engine/trainer.py

@@ -0,0 +1,258 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+
+import logging
+import os
+from collections import OrderedDict
+from typing import List, Optional, Union
+import torch
+from torch import nn
+
+from detectron2.checkpoint import DetectionCheckpointer
+from detectron2.config import CfgNode
+from detectron2.engine import DefaultTrainer
+from detectron2.evaluation import (
+    DatasetEvaluator,
+    DatasetEvaluators,
+    inference_on_dataset,
+    print_csv_format,
+)
+from detectron2.solver.build import get_default_optimizer_params, maybe_add_gradient_clipping
+from detectron2.utils import comm
+from detectron2.utils.events import EventWriter, get_event_storage
+
+from densepose import DensePoseDatasetMapperTTA, DensePoseGeneralizedRCNNWithTTA, load_from_cfg
+from densepose.data import (
+    DatasetMapper,
+    build_combined_loader,
+    build_detection_test_loader,
+    build_detection_train_loader,
+    build_inference_based_loaders,
+    has_inference_based_loaders,
+)
+from densepose.evaluation.d2_evaluator_adapter import Detectron2COCOEvaluatorAdapter
+from densepose.evaluation.evaluator import DensePoseCOCOEvaluator, build_densepose_evaluator_storage
+from densepose.modeling.cse import Embedder
+
+
+class SampleCountingLoader:
+    def __init__(self, loader):
+        self.loader = loader
+
+    def __iter__(self):
+        it = iter(self.loader)
+        storage = get_event_storage()
+        while True:
+            try:
+                batch = next(it)
+                num_inst_per_dataset = {}
+                for data in batch:
+                    dataset_name = data["dataset"]
+                    if dataset_name not in num_inst_per_dataset:
+                        num_inst_per_dataset[dataset_name] = 0
+                    num_inst = len(data["instances"])
+                    num_inst_per_dataset[dataset_name] += num_inst
+                for dataset_name in num_inst_per_dataset:
+                    storage.put_scalar(f"batch/{dataset_name}", num_inst_per_dataset[dataset_name])
+                yield batch
+            except StopIteration:
+                break
+
+
+class SampleCountMetricPrinter(EventWriter):
+    def __init__(self):
+        self.logger = logging.getLogger(__name__)
+
+    def write(self):
+        storage = get_event_storage()
+        batch_stats_strs = []
+        for key, buf in storage.histories().items():
+            if key.startswith("batch/"):
+                batch_stats_strs.append(f"{key} {buf.avg(20)}")
+        self.logger.info(", ".join(batch_stats_strs))
+
+
+class Trainer(DefaultTrainer):
+    @classmethod
+    def extract_embedder_from_model(cls, model: nn.Module) -> Optional[Embedder]:
+        if isinstance(model, nn.parallel.DistributedDataParallel):
+            model = model.module
+        if hasattr(model, "roi_heads") and hasattr(model.roi_heads, "embedder"):
+            return model.roi_heads.embedder
+        return None
+
+    # TODO: the only reason to copy the base class code here is to pass the embedder from
+    # the model to the evaluator; that should be refactored to avoid unnecessary copy-pasting
+    @classmethod
+    def test(
+        cls,
+        cfg: CfgNode,
+        model: nn.Module,
+        evaluators: Optional[Union[DatasetEvaluator, List[DatasetEvaluator]]] = None,
+    ):
+        """
+        Args:
+            cfg (CfgNode):
+            model (nn.Module):
+            evaluators (DatasetEvaluator, list[DatasetEvaluator] or None): if None, will call
+                :meth:`build_evaluator`. Otherwise, must have the same length as
+                ``cfg.DATASETS.TEST``.
+
+        Returns:
+            dict: a dict of result metrics
+        """
+        logger = logging.getLogger(__name__)
+        if isinstance(evaluators, DatasetEvaluator):
+            evaluators = [evaluators]
+        if evaluators is not None:
+            assert len(cfg.DATASETS.TEST) == len(evaluators), "{} != {}".format(
+                len(cfg.DATASETS.TEST), len(evaluators)
+            )
+
+        results = OrderedDict()
+        for idx, dataset_name in enumerate(cfg.DATASETS.TEST):
+            data_loader = cls.build_test_loader(cfg, dataset_name)
+            # When evaluators are passed in as arguments,
+            # implicitly assume that evaluators can be created before data_loader.
+            if evaluators is not None:
+                evaluator = evaluators[idx]
+            else:
+                try:
+                    embedder = cls.extract_embedder_from_model(model)
+                    evaluator = cls.build_evaluator(cfg, dataset_name, embedder=embedder)
+                except NotImplementedError:
+                    logger.warn(
+                        "No evaluator found. Use `DefaultTrainer.test(evaluators=)`, "
+                        "or implement its `build_evaluator` method."
+                    )
+                    results[dataset_name] = {}
+                    continue
+            if cfg.DENSEPOSE_EVALUATION.DISTRIBUTED_INFERENCE or comm.is_main_process():
+                results_i = inference_on_dataset(model, data_loader, evaluator)
+            else:
+                results_i = {}
+            results[dataset_name] = results_i
+            if comm.is_main_process():
+                assert isinstance(
+                    results_i, dict
+                ), "Evaluator must return a dict on the main process. Got {} instead.".format(
+                    results_i
+                )
+                logger.info("Evaluation results for {} in csv format:".format(dataset_name))
+                print_csv_format(results_i)
+
+        if len(results) == 1:
+            results = list(results.values())[0]
+        return results
+
+    @classmethod
+    def build_evaluator(
+        cls,
+        cfg: CfgNode,
+        dataset_name: str,
+        output_folder: Optional[str] = None,
+        embedder: Optional[Embedder] = None,
+    ) -> DatasetEvaluators:
+        if output_folder is None:
+            output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
+        evaluators = []
+        distributed = cfg.DENSEPOSE_EVALUATION.DISTRIBUTED_INFERENCE
+        # Note: we currently use COCO evaluator for both COCO and LVIS datasets
+        # to have compatible metrics. LVIS bbox evaluator could also be used
+        # with an adapter to properly handle filtered / mapped categories
+        # evaluator_type = MetadataCatalog.get(dataset_name).evaluator_type
+        # if evaluator_type == "coco":
+        #     evaluators.append(COCOEvaluator(dataset_name, output_dir=output_folder))
+        # elif evaluator_type == "lvis":
+        #     evaluators.append(LVISEvaluator(dataset_name, output_dir=output_folder))
+        evaluators.append(
+            Detectron2COCOEvaluatorAdapter(
+                dataset_name, output_dir=output_folder, distributed=distributed
+            )
+        )
+        if cfg.MODEL.DENSEPOSE_ON:
+            storage = build_densepose_evaluator_storage(cfg, output_folder)
+            evaluators.append(
+                DensePoseCOCOEvaluator(
+                    dataset_name,
+                    distributed,
+                    output_folder,
+                    evaluator_type=cfg.DENSEPOSE_EVALUATION.TYPE,
+                    min_iou_threshold=cfg.DENSEPOSE_EVALUATION.MIN_IOU_THRESHOLD,
+                    storage=storage,
+                    embedder=embedder,
+                    should_evaluate_mesh_alignment=cfg.DENSEPOSE_EVALUATION.EVALUATE_MESH_ALIGNMENT,
+                    mesh_alignment_mesh_names=cfg.DENSEPOSE_EVALUATION.MESH_ALIGNMENT_MESH_NAMES,
+                )
+            )
+        return DatasetEvaluators(evaluators)
+
+    @classmethod
+    def build_optimizer(cls, cfg: CfgNode, model: nn.Module):
+        params = get_default_optimizer_params(
+            model,
+            base_lr=cfg.SOLVER.BASE_LR,
+            weight_decay_norm=cfg.SOLVER.WEIGHT_DECAY_NORM,
+            bias_lr_factor=cfg.SOLVER.BIAS_LR_FACTOR,
+            weight_decay_bias=cfg.SOLVER.WEIGHT_DECAY_BIAS,
+            overrides={
+                "features": {
+                    "lr": cfg.SOLVER.BASE_LR * cfg.MODEL.ROI_DENSEPOSE_HEAD.CSE.FEATURES_LR_FACTOR,
+                },
+                "embeddings": {
+                    "lr": cfg.SOLVER.BASE_LR * cfg.MODEL.ROI_DENSEPOSE_HEAD.CSE.EMBEDDING_LR_FACTOR,
+                },
+            },
+        )
+        optimizer = torch.optim.SGD(
+            params,
+            cfg.SOLVER.BASE_LR,
+            momentum=cfg.SOLVER.MOMENTUM,
+            nesterov=cfg.SOLVER.NESTEROV,
+            weight_decay=cfg.SOLVER.WEIGHT_DECAY,
+        )
+        # pyre-fixme[6]: For 2nd param expected `Type[Optimizer]` but got `SGD`.
+        return maybe_add_gradient_clipping(cfg, optimizer)
+
+    @classmethod
+    def build_test_loader(cls, cfg: CfgNode, dataset_name):
+        return build_detection_test_loader(cfg, dataset_name, mapper=DatasetMapper(cfg, False))
+
+    @classmethod
+    def build_train_loader(cls, cfg: CfgNode):
+        data_loader = build_detection_train_loader(cfg, mapper=DatasetMapper(cfg, True))
+        if not has_inference_based_loaders(cfg):
+            return data_loader
+        model = cls.build_model(cfg)
+        model.to(cfg.BOOTSTRAP_MODEL.DEVICE)
+        DetectionCheckpointer(model).resume_or_load(cfg.BOOTSTRAP_MODEL.WEIGHTS, resume=False)
+        inference_based_loaders, ratios = build_inference_based_loaders(cfg, model)
+        loaders = [data_loader] + inference_based_loaders
+        ratios = [1.0] + ratios
+        combined_data_loader = build_combined_loader(cfg, loaders, ratios)
+        sample_counting_loader = SampleCountingLoader(combined_data_loader)
+        return sample_counting_loader
+
+    def build_writers(self):
+        writers = super().build_writers()
+        writers.append(SampleCountMetricPrinter())
+        return writers
+
+    @classmethod
+    def test_with_TTA(cls, cfg: CfgNode, model):
+        logger = logging.getLogger("detectron2.trainer")
+        # In the end of training, run an evaluation with TTA
+        # Only support some R-CNN models.
+        logger.info("Running inference with test-time augmentation ...")
+        transform_data = load_from_cfg(cfg)
+        model = DensePoseGeneralizedRCNNWithTTA(
+            cfg, model, transform_data, DensePoseDatasetMapperTTA(cfg)
+        )
+        evaluators = [
+            cls.build_evaluator(
+                cfg, name, output_folder=os.path.join(cfg.OUTPUT_DIR, "inference_TTA")
+            )
+            for name in cfg.DATASETS.TEST
+        ]
+        res = cls.test(cfg, model, evaluators)  # pyre-ignore[6]
+        res = OrderedDict({k + "_TTA": v for k, v in res.items()})
+        return res

3rdparty/densepose/evaluation/__init__.py

@@ -0,0 +1,3 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from .evaluator import DensePoseCOCOEvaluator

3rdparty/densepose/evaluation/d2_evaluator_adapter.py

@@ -0,0 +1,50 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from detectron2.data.catalog import Metadata
+from detectron2.evaluation import COCOEvaluator
+
+from densepose.data.datasets.coco import (
+    get_contiguous_id_to_category_id_map,
+    maybe_filter_categories_cocoapi,
+)
+
+
+def _maybe_add_iscrowd_annotations(cocoapi) -> None:
+    for ann in cocoapi.dataset["annotations"]:
+        if "iscrowd" not in ann:
+            ann["iscrowd"] = 0
+
+
+class Detectron2COCOEvaluatorAdapter(COCOEvaluator):
+    def __init__(
+        self,
+        dataset_name,
+        output_dir=None,
+        distributed=True,
+    ):
+        super().__init__(dataset_name, output_dir=output_dir, distributed=distributed)
+        maybe_filter_categories_cocoapi(dataset_name, self._coco_api)
+        _maybe_add_iscrowd_annotations(self._coco_api)
+        # substitute category metadata to account for categories
+        # that are mapped to the same contiguous id
+        if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
+            self._maybe_substitute_metadata()
+
+    def _maybe_substitute_metadata(self):
+        cont_id_2_cat_id = get_contiguous_id_to_category_id_map(self._metadata)
+        cat_id_2_cont_id = self._metadata.thing_dataset_id_to_contiguous_id
+        if len(cont_id_2_cat_id) == len(cat_id_2_cont_id):
+            return
+
+        cat_id_2_cont_id_injective = {}
+        for cat_id, cont_id in cat_id_2_cont_id.items():
+            if (cont_id in cont_id_2_cat_id) and (cont_id_2_cat_id[cont_id] == cat_id):
+                cat_id_2_cont_id_injective[cat_id] = cont_id
+
+        metadata_new = Metadata(name=self._metadata.name)
+        for key, value in self._metadata.__dict__.items():
+            if key == "thing_dataset_id_to_contiguous_id":
+                setattr(metadata_new, key, cat_id_2_cont_id_injective)
+            else:
+                setattr(metadata_new, key, value)
+        self._metadata = metadata_new

3rdparty/densepose/evaluation/densepose_coco_evaluation.py

@@ -0,0 +1,1303 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+# This is a modified version of cocoeval.py where we also have the densepose evaluation.
+
+__author__ = "tsungyi"
+
+import copy
+import datetime
+import logging
+import numpy as np
+import pickle
+import time
+from collections import defaultdict
+from enum import Enum
+from typing import Any, Dict, Tuple
+import scipy.spatial.distance as ssd
+import torch
+import torch.nn.functional as F
+from pycocotools import mask as maskUtils
+from scipy.io import loadmat
+from scipy.ndimage import zoom as spzoom
+
+from detectron2.utils.file_io import PathManager
+
+from densepose.converters.chart_output_to_chart_result import resample_uv_tensors_to_bbox
+from densepose.converters.segm_to_mask import (
+    resample_coarse_segm_tensor_to_bbox,
+    resample_fine_and_coarse_segm_tensors_to_bbox,
+)
+from densepose.modeling.cse.utils import squared_euclidean_distance_matrix
+from densepose.structures import DensePoseDataRelative
+from densepose.structures.mesh import create_mesh
+
+logger = logging.getLogger(__name__)
+
+
+class DensePoseEvalMode(str, Enum):
+    # use both masks and geodesic distances (GPS * IOU) to compute scores
+    GPSM = "gpsm"
+    # use only geodesic distances (GPS)  to compute scores
+    GPS = "gps"
+    # use only masks (IOU) to compute scores
+    IOU = "iou"
+
+
+class DensePoseDataMode(str, Enum):
+    # use estimated IUV data (default mode)
+    IUV_DT = "iuvdt"
+    # use ground truth IUV data
+    IUV_GT = "iuvgt"
+    # use ground truth labels I and set UV to 0
+    I_GT_UV_0 = "igtuv0"
+    # use ground truth labels I and estimated UV coordinates
+    I_GT_UV_DT = "igtuvdt"
+    # use estimated labels I and set UV to 0
+    I_DT_UV_0 = "idtuv0"
+
+
+class DensePoseCocoEval:
+    # Interface for evaluating detection on the Microsoft COCO dataset.
+    #
+    # The usage for CocoEval is as follows:
+    #  cocoGt=..., cocoDt=...       # load dataset and results
+    #  E = CocoEval(cocoGt,cocoDt); # initialize CocoEval object
+    #  E.params.recThrs = ...;      # set parameters as desired
+    #  E.evaluate();                # run per image evaluation
+    #  E.accumulate();              # accumulate per image results
+    #  E.summarize();               # display summary metrics of results
+    # For example usage see evalDemo.m and http://mscoco.org/.
+    #
+    # The evaluation parameters are as follows (defaults in brackets):
+    #  imgIds     - [all] N img ids to use for evaluation
+    #  catIds     - [all] K cat ids to use for evaluation
+    #  iouThrs    - [.5:.05:.95] T=10 IoU thresholds for evaluation
+    #  recThrs    - [0:.01:1] R=101 recall thresholds for evaluation
+    #  areaRng    - [...] A=4 object area ranges for evaluation
+    #  maxDets    - [1 10 100] M=3 thresholds on max detections per image
+    #  iouType    - ['segm'] set iouType to 'segm', 'bbox', 'keypoints' or 'densepose'
+    #  iouType replaced the now DEPRECATED useSegm parameter.
+    #  useCats    - [1] if true use category labels for evaluation
+    # Note: if useCats=0 category labels are ignored as in proposal scoring.
+    # Note: multiple areaRngs [Ax2] and maxDets [Mx1] can be specified.
+    #
+    # evaluate(): evaluates detections on every image and every category and
+    # concats the results into the "evalImgs" with fields:
+    #  dtIds      - [1xD] id for each of the D detections (dt)
+    #  gtIds      - [1xG] id for each of the G ground truths (gt)
+    #  dtMatches  - [TxD] matching gt id at each IoU or 0
+    #  gtMatches  - [TxG] matching dt id at each IoU or 0
+    #  dtScores   - [1xD] confidence of each dt
+    #  gtIgnore   - [1xG] ignore flag for each gt
+    #  dtIgnore   - [TxD] ignore flag for each dt at each IoU
+    #
+    # accumulate(): accumulates the per-image, per-category evaluation
+    # results in "evalImgs" into the dictionary "eval" with fields:
+    #  params     - parameters used for evaluation
+    #  date       - date evaluation was performed
+    #  counts     - [T,R,K,A,M] parameter dimensions (see above)
+    #  precision  - [TxRxKxAxM] precision for every evaluation setting
+    #  recall     - [TxKxAxM] max recall for every evaluation setting
+    # Note: precision and recall==-1 for settings with no gt objects.
+    #
+    # See also coco, mask, pycocoDemo, pycocoEvalDemo
+    #
+    # Microsoft COCO Toolbox.      version 2.0
+    # Data, paper, and tutorials available at:  http://mscoco.org/
+    # Code written by Piotr Dollar and Tsung-Yi Lin, 2015.
+    # Licensed under the Simplified BSD License [see coco/license.txt]
+    def __init__(
+        self,
+        cocoGt=None,
+        cocoDt=None,
+        iouType: str = "densepose",
+        multi_storage=None,
+        embedder=None,
+        dpEvalMode: DensePoseEvalMode = DensePoseEvalMode.GPS,
+        dpDataMode: DensePoseDataMode = DensePoseDataMode.IUV_DT,
+    ):
+        """
+        Initialize CocoEval using coco APIs for gt and dt
+        :param cocoGt: coco object with ground truth annotations
+        :param cocoDt: coco object with detection results
+        :return: None
+        """
+        self.cocoGt = cocoGt  # ground truth COCO API
+        self.cocoDt = cocoDt  # detections COCO API
+        self.multi_storage = multi_storage
+        self.embedder = embedder
+        self._dpEvalMode = dpEvalMode
+        self._dpDataMode = dpDataMode
+        self.evalImgs = defaultdict(list)  # per-image per-category eval results [KxAxI]
+        self.eval = {}  # accumulated evaluation results
+        self._gts = defaultdict(list)  # gt for evaluation
+        self._dts = defaultdict(list)  # dt for evaluation
+        self.params = Params(iouType=iouType)  # parameters
+        self._paramsEval = {}  # parameters for evaluation
+        self.stats = []  # result summarization
+        self.ious = {}  # ious between all gts and dts
+        if cocoGt is not None:
+            self.params.imgIds = sorted(cocoGt.getImgIds())
+            self.params.catIds = sorted(cocoGt.getCatIds())
+        self.ignoreThrBB = 0.7
+        self.ignoreThrUV = 0.9
+
+    def _loadGEval(self):
+        smpl_subdiv_fpath = PathManager.get_local_path(
+            "https://dl.fbaipublicfiles.com/densepose/data/SMPL_subdiv.mat"
+        )
+        pdist_transform_fpath = PathManager.get_local_path(
+            "https://dl.fbaipublicfiles.com/densepose/data/SMPL_SUBDIV_TRANSFORM.mat"
+        )
+        pdist_matrix_fpath = PathManager.get_local_path(
+            "https://dl.fbaipublicfiles.com/densepose/data/Pdist_matrix.pkl", timeout_sec=120
+        )
+        SMPL_subdiv = loadmat(smpl_subdiv_fpath)
+        self.PDIST_transform = loadmat(pdist_transform_fpath)
+        self.PDIST_transform = self.PDIST_transform["index"].squeeze()
+        UV = np.array([SMPL_subdiv["U_subdiv"], SMPL_subdiv["V_subdiv"]]).squeeze()
+        ClosestVertInds = np.arange(UV.shape[1]) + 1
+        self.Part_UVs = []
+        self.Part_ClosestVertInds = []
+        for i in np.arange(24):
+            self.Part_UVs.append(UV[:, SMPL_subdiv["Part_ID_subdiv"].squeeze() == (i + 1)])
+            self.Part_ClosestVertInds.append(
+                ClosestVertInds[SMPL_subdiv["Part_ID_subdiv"].squeeze() == (i + 1)]
+            )
+
+        with open(pdist_matrix_fpath, "rb") as hFile:
+            arrays = pickle.load(hFile, encoding="latin1")
+        self.Pdist_matrix = arrays["Pdist_matrix"]
+        self.Part_ids = np.array(SMPL_subdiv["Part_ID_subdiv"].squeeze())
+        # Mean geodesic distances for parts.
+        self.Mean_Distances = np.array([0, 0.351, 0.107, 0.126, 0.237, 0.173, 0.142, 0.128, 0.150])
+        # Coarse Part labels.
+        self.CoarseParts = np.array(
+            [0, 1, 1, 2, 2, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6, 6, 6, 7, 7, 7, 7, 8, 8]
+        )
+
+    def _prepare(self):
+        """
+        Prepare ._gts and ._dts for evaluation based on params
+        :return: None
+        """
+
+        def _toMask(anns, coco):
+            # modify ann['segmentation'] by reference
+            for ann in anns:
+                # safeguard for invalid segmentation annotation;
+                # annotations containing empty lists exist in the posetrack
+                # dataset. This is not a correct segmentation annotation
+                # in terms of COCO format; we need to deal with it somehow
+                segm = ann["segmentation"]
+                if type(segm) == list and len(segm) == 0:
+                    ann["segmentation"] = None
+                    continue
+                rle = coco.annToRLE(ann)
+                ann["segmentation"] = rle
+
+        def _getIgnoreRegion(iid, coco):
+            img = coco.imgs[iid]
+
+            if "ignore_regions_x" not in img.keys():
+                return None
+
+            if len(img["ignore_regions_x"]) == 0:
+                return None
+
+            rgns_merged = [
+                [v for xy in zip(region_x, region_y) for v in xy]
+                for region_x, region_y in zip(img["ignore_regions_x"], img["ignore_regions_y"])
+            ]
+            rles = maskUtils.frPyObjects(rgns_merged, img["height"], img["width"])
+            rle = maskUtils.merge(rles)
+            return maskUtils.decode(rle)
+
+        def _checkIgnore(dt, iregion):
+            if iregion is None:
+                return True
+
+            bb = np.array(dt["bbox"]).astype(int)
+            x1, y1, x2, y2 = bb[0], bb[1], bb[0] + bb[2], bb[1] + bb[3]
+            x2 = min([x2, iregion.shape[1]])
+            y2 = min([y2, iregion.shape[0]])
+
+            if bb[2] * bb[3] == 0:
+                return False
+
+            crop_iregion = iregion[y1:y2, x1:x2]
+
+            if crop_iregion.sum() == 0:
+                return True
+
+            if "densepose" not in dt.keys():  # filtering boxes
+                return crop_iregion.sum() / bb[2] / bb[3] < self.ignoreThrBB
+
+            # filtering UVs
+            ignoremask = np.require(crop_iregion, requirements=["F"])
+            mask = self._extract_mask(dt)
+            uvmask = np.require(np.asarray(mask > 0), dtype=np.uint8, requirements=["F"])
+            uvmask_ = maskUtils.encode(uvmask)
+            ignoremask_ = maskUtils.encode(ignoremask)
+            uviou = maskUtils.iou([uvmask_], [ignoremask_], [1])[0]
+            return uviou < self.ignoreThrUV
+
+        p = self.params
+
+        if p.useCats:
+            gts = self.cocoGt.loadAnns(self.cocoGt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
+            dts = self.cocoDt.loadAnns(self.cocoDt.getAnnIds(imgIds=p.imgIds, catIds=p.catIds))
+        else:
+            gts = self.cocoGt.loadAnns(self.cocoGt.getAnnIds(imgIds=p.imgIds))
+            dts = self.cocoDt.loadAnns(self.cocoDt.getAnnIds(imgIds=p.imgIds))
+
+        imns = self.cocoGt.loadImgs(p.imgIds)
+        self.size_mapping = {}
+        for im in imns:
+            self.size_mapping[im["id"]] = [im["height"], im["width"]]
+
+        # if iouType == 'uv', add point gt annotations
+        if p.iouType == "densepose":
+            self._loadGEval()
+
+        # convert ground truth to mask if iouType == 'segm'
+        if p.iouType == "segm":
+            _toMask(gts, self.cocoGt)
+            _toMask(dts, self.cocoDt)
+
+        # set ignore flag
+        for gt in gts:
+            gt["ignore"] = gt["ignore"] if "ignore" in gt else 0
+            gt["ignore"] = "iscrowd" in gt and gt["iscrowd"]
+            if p.iouType == "keypoints":
+                gt["ignore"] = (gt["num_keypoints"] == 0) or gt["ignore"]
+            if p.iouType == "densepose":
+                gt["ignore"] = ("dp_x" in gt) == 0
+            if p.iouType == "segm":
+                gt["ignore"] = gt["segmentation"] is None
+
+        self._gts = defaultdict(list)  # gt for evaluation
+        self._dts = defaultdict(list)  # dt for evaluation
+        self._igrgns = defaultdict(list)
+
+        for gt in gts:
+            iid = gt["image_id"]
+            if iid not in self._igrgns.keys():
+                self._igrgns[iid] = _getIgnoreRegion(iid, self.cocoGt)
+            if _checkIgnore(gt, self._igrgns[iid]):
+                self._gts[iid, gt["category_id"]].append(gt)
+        for dt in dts:
+            iid = dt["image_id"]
+            if (iid not in self._igrgns) or _checkIgnore(dt, self._igrgns[iid]):
+                self._dts[iid, dt["category_id"]].append(dt)
+
+        self.evalImgs = defaultdict(list)  # per-image per-category evaluation results
+        self.eval = {}  # accumulated evaluation results
+
+    def evaluate(self):
+        """
+        Run per image evaluation on given images and store results (a list of dict) in self.evalImgs
+        :return: None
+        """
+        tic = time.time()
+        logger.info("Running per image DensePose evaluation... {}".format(self.params.iouType))
+        p = self.params
+        # add backward compatibility if useSegm is specified in params
+        if p.useSegm is not None:
+            p.iouType = "segm" if p.useSegm == 1 else "bbox"
+            logger.info("useSegm (deprecated) is not None. Running DensePose evaluation")
+        p.imgIds = list(np.unique(p.imgIds))
+        if p.useCats:
+            p.catIds = list(np.unique(p.catIds))
+        p.maxDets = sorted(p.maxDets)
+        self.params = p
+
+        self._prepare()
+        # loop through images, area range, max detection number
+        catIds = p.catIds if p.useCats else [-1]
+
+        if p.iouType in ["segm", "bbox"]:
+            computeIoU = self.computeIoU
+        elif p.iouType == "keypoints":
+            computeIoU = self.computeOks
+        elif p.iouType == "densepose":
+            computeIoU = self.computeOgps
+            if self._dpEvalMode in {DensePoseEvalMode.GPSM, DensePoseEvalMode.IOU}:
+                self.real_ious = {
+                    (imgId, catId): self.computeDPIoU(imgId, catId)
+                    for imgId in p.imgIds
+                    for catId in catIds
+                }
+
+        self.ious = {
+            (imgId, catId): computeIoU(imgId, catId) for imgId in p.imgIds for catId in catIds
+        }
+
+        evaluateImg = self.evaluateImg
+        maxDet = p.maxDets[-1]
+        self.evalImgs = [
+            evaluateImg(imgId, catId, areaRng, maxDet)
+            for catId in catIds
+            for areaRng in p.areaRng
+            for imgId in p.imgIds
+        ]
+        self._paramsEval = copy.deepcopy(self.params)
+        toc = time.time()
+        logger.info("DensePose evaluation DONE (t={:0.2f}s).".format(toc - tic))
+
+    def getDensePoseMask(self, polys):
+        maskGen = np.zeros([256, 256])
+        stop = min(len(polys) + 1, 15)
+        for i in range(1, stop):
+            if polys[i - 1]:
+                currentMask = maskUtils.decode(polys[i - 1])
+                maskGen[currentMask > 0] = i
+        return maskGen
+
+    def _generate_rlemask_on_image(self, mask, imgId, data):
+        bbox_xywh = np.array(data["bbox"])
+        x, y, w, h = bbox_xywh
+        im_h, im_w = self.size_mapping[imgId]
+        im_mask = np.zeros((im_h, im_w), dtype=np.uint8)
+        if mask is not None:
+            x0 = max(int(x), 0)
+            x1 = min(int(x + w), im_w, int(x) + mask.shape[1])
+            y0 = max(int(y), 0)
+            y1 = min(int(y + h), im_h, int(y) + mask.shape[0])
+            y = int(y)
+            x = int(x)
+            im_mask[y0:y1, x0:x1] = mask[y0 - y : y1 - y, x0 - x : x1 - x]
+        im_mask = np.require(np.asarray(im_mask > 0), dtype=np.uint8, requirements=["F"])
+        rle_mask = maskUtils.encode(np.array(im_mask[:, :, np.newaxis], order="F"))[0]
+        return rle_mask
+
+    def computeDPIoU(self, imgId, catId):
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return []
+        inds = np.argsort([-d["score"] for d in dt], kind="mergesort")
+        dt = [dt[i] for i in inds]
+        if len(dt) > p.maxDets[-1]:
+            dt = dt[0 : p.maxDets[-1]]
+
+        gtmasks = []
+        for g in gt:
+            if DensePoseDataRelative.S_KEY in g:
+                # convert DensePose mask to a binary mask
+                mask = np.minimum(self.getDensePoseMask(g[DensePoseDataRelative.S_KEY]), 1.0)
+                _, _, w, h = g["bbox"]
+                scale_x = float(max(w, 1)) / mask.shape[1]
+                scale_y = float(max(h, 1)) / mask.shape[0]
+                mask = spzoom(mask, (scale_y, scale_x), order=1, prefilter=False)
+                mask = np.array(mask > 0.5, dtype=np.uint8)
+                rle_mask = self._generate_rlemask_on_image(mask, imgId, g)
+            elif "segmentation" in g:
+                segmentation = g["segmentation"]
+                if isinstance(segmentation, list) and segmentation:
+                    # polygons
+                    im_h, im_w = self.size_mapping[imgId]
+                    rles = maskUtils.frPyObjects(segmentation, im_h, im_w)
+                    rle_mask = maskUtils.merge(rles)
+                elif isinstance(segmentation, dict):
+                    if isinstance(segmentation["counts"], list):
+                        # uncompressed RLE
+                        im_h, im_w = self.size_mapping[imgId]
+                        rle_mask = maskUtils.frPyObjects(segmentation, im_h, im_w)
+                    else:
+                        # compressed RLE
+                        rle_mask = segmentation
+                else:
+                    rle_mask = self._generate_rlemask_on_image(None, imgId, g)
+            else:
+                rle_mask = self._generate_rlemask_on_image(None, imgId, g)
+            gtmasks.append(rle_mask)
+
+        dtmasks = []
+        for d in dt:
+            mask = self._extract_mask(d)
+            mask = np.require(np.asarray(mask > 0), dtype=np.uint8, requirements=["F"])
+            rle_mask = self._generate_rlemask_on_image(mask, imgId, d)
+            dtmasks.append(rle_mask)
+
+        # compute iou between each dt and gt region
+        iscrowd = [int(o.get("iscrowd", 0)) for o in gt]
+        iousDP = maskUtils.iou(dtmasks, gtmasks, iscrowd)
+        return iousDP
+
+    def computeIoU(self, imgId, catId):
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return []
+        inds = np.argsort([-d["score"] for d in dt], kind="mergesort")
+        dt = [dt[i] for i in inds]
+        if len(dt) > p.maxDets[-1]:
+            dt = dt[0 : p.maxDets[-1]]
+
+        if p.iouType == "segm":
+            g = [g["segmentation"] for g in gt if g["segmentation"] is not None]
+            d = [d["segmentation"] for d in dt if d["segmentation"] is not None]
+        elif p.iouType == "bbox":
+            g = [g["bbox"] for g in gt]
+            d = [d["bbox"] for d in dt]
+        else:
+            raise Exception("unknown iouType for iou computation")
+
+        # compute iou between each dt and gt region
+        iscrowd = [int(o.get("iscrowd", 0)) for o in gt]
+        ious = maskUtils.iou(d, g, iscrowd)
+        return ious
+
+    def computeOks(self, imgId, catId):
+        p = self.params
+        # dimension here should be Nxm
+        gts = self._gts[imgId, catId]
+        dts = self._dts[imgId, catId]
+        inds = np.argsort([-d["score"] for d in dts], kind="mergesort")
+        dts = [dts[i] for i in inds]
+        if len(dts) > p.maxDets[-1]:
+            dts = dts[0 : p.maxDets[-1]]
+        # if len(gts) == 0 and len(dts) == 0:
+        if len(gts) == 0 or len(dts) == 0:
+            return []
+        ious = np.zeros((len(dts), len(gts)))
+        sigmas = (
+            np.array(
+                [
+                    0.26,
+                    0.25,
+                    0.25,
+                    0.35,
+                    0.35,
+                    0.79,
+                    0.79,
+                    0.72,
+                    0.72,
+                    0.62,
+                    0.62,
+                    1.07,
+                    1.07,
+                    0.87,
+                    0.87,
+                    0.89,
+                    0.89,
+                ]
+            )
+            / 10.0
+        )
+        vars = (sigmas * 2) ** 2
+        k = len(sigmas)
+        # compute oks between each detection and ground truth object
+        for j, gt in enumerate(gts):
+            # create bounds for ignore regions(double the gt bbox)
+            g = np.array(gt["keypoints"])
+            xg = g[0::3]
+            yg = g[1::3]
+            vg = g[2::3]
+            k1 = np.count_nonzero(vg > 0)
+            bb = gt["bbox"]
+            x0 = bb[0] - bb[2]
+            x1 = bb[0] + bb[2] * 2
+            y0 = bb[1] - bb[3]
+            y1 = bb[1] + bb[3] * 2
+            for i, dt in enumerate(dts):
+                d = np.array(dt["keypoints"])
+                xd = d[0::3]
+                yd = d[1::3]
+                if k1 > 0:
+                    # measure the per-keypoint distance if keypoints visible
+                    dx = xd - xg
+                    dy = yd - yg
+                else:
+                    # measure minimum distance to keypoints in (x0,y0) & (x1,y1)
+                    z = np.zeros(k)
+                    dx = np.max((z, x0 - xd), axis=0) + np.max((z, xd - x1), axis=0)
+                    dy = np.max((z, y0 - yd), axis=0) + np.max((z, yd - y1), axis=0)
+                e = (dx**2 + dy**2) / vars / (gt["area"] + np.spacing(1)) / 2
+                if k1 > 0:
+                    e = e[vg > 0]
+                ious[i, j] = np.sum(np.exp(-e)) / e.shape[0]
+        return ious
+
+    def _extract_mask(self, dt: Dict[str, Any]) -> np.ndarray:
+        if "densepose" in dt:
+            densepose_results_quantized = dt["densepose"]
+            return densepose_results_quantized.labels_uv_uint8[0].numpy()
+        elif "cse_mask" in dt:
+            return dt["cse_mask"]
+        elif "coarse_segm" in dt:
+            dy = max(int(dt["bbox"][3]), 1)
+            dx = max(int(dt["bbox"][2]), 1)
+            return (
+                F.interpolate(
+                    dt["coarse_segm"].unsqueeze(0),
+                    (dy, dx),
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                .squeeze(0)
+                .argmax(0)
+                .numpy()
+                .astype(np.uint8)
+            )
+        elif "record_id" in dt:
+            assert (
+                self.multi_storage is not None
+            ), f"Storage record id encountered in a detection {dt}, but no storage provided!"
+            record = self.multi_storage.get(dt["rank"], dt["record_id"])
+            coarse_segm = record["coarse_segm"]
+            dy = max(int(dt["bbox"][3]), 1)
+            dx = max(int(dt["bbox"][2]), 1)
+            return (
+                F.interpolate(
+                    coarse_segm.unsqueeze(0),
+                    (dy, dx),
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                .squeeze(0)
+                .argmax(0)
+                .numpy()
+                .astype(np.uint8)
+            )
+        else:
+            raise Exception(f"No mask data in the detection: {dt}")
+        raise ValueError('The prediction dict needs to contain either "densepose" or "cse_mask"')
+
+    def _extract_iuv(
+        self, densepose_data: np.ndarray, py: np.ndarray, px: np.ndarray, gt: Dict[str, Any]
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Extract arrays of I, U and V values at given points as numpy arrays
+        given the data mode stored in self._dpDataMode
+        """
+        if self._dpDataMode == DensePoseDataMode.IUV_DT:
+            # estimated labels and UV (default)
+            ipoints = densepose_data[0, py, px]
+            upoints = densepose_data[1, py, px] / 255.0  # convert from uint8 by /255.
+            vpoints = densepose_data[2, py, px] / 255.0
+        elif self._dpDataMode == DensePoseDataMode.IUV_GT:
+            # ground truth
+            ipoints = np.array(gt["dp_I"])
+            upoints = np.array(gt["dp_U"])
+            vpoints = np.array(gt["dp_V"])
+        elif self._dpDataMode == DensePoseDataMode.I_GT_UV_0:
+            # ground truth labels, UV = 0
+            ipoints = np.array(gt["dp_I"])
+            upoints = upoints * 0.0
+            vpoints = vpoints * 0.0
+        elif self._dpDataMode == DensePoseDataMode.I_GT_UV_DT:
+            # ground truth labels, estimated UV
+            ipoints = np.array(gt["dp_I"])
+            upoints = densepose_data[1, py, px] / 255.0  # convert from uint8 by /255.
+            vpoints = densepose_data[2, py, px] / 255.0
+        elif self._dpDataMode == DensePoseDataMode.I_DT_UV_0:
+            # estimated labels, UV = 0
+            ipoints = densepose_data[0, py, px]
+            upoints = upoints * 0.0
+            vpoints = vpoints * 0.0
+        else:
+            raise ValueError(f"Unknown data mode: {self._dpDataMode}")
+        return ipoints, upoints, vpoints
+
+    def computeOgps_single_pair(self, dt, gt, py, px, pt_mask):
+        if "densepose" in dt:
+            ipoints, upoints, vpoints = self.extract_iuv_from_quantized(dt, gt, py, px, pt_mask)
+            return self.computeOgps_single_pair_iuv(dt, gt, ipoints, upoints, vpoints)
+        elif "u" in dt:
+            ipoints, upoints, vpoints = self.extract_iuv_from_raw(dt, gt, py, px, pt_mask)
+            return self.computeOgps_single_pair_iuv(dt, gt, ipoints, upoints, vpoints)
+        elif "record_id" in dt:
+            assert (
+                self.multi_storage is not None
+            ), f"Storage record id encountered in detection {dt}, but no storage provided!"
+            record = self.multi_storage.get(dt["rank"], dt["record_id"])
+            record["bbox"] = dt["bbox"]
+            if "u" in record:
+                ipoints, upoints, vpoints = self.extract_iuv_from_raw(record, gt, py, px, pt_mask)
+                return self.computeOgps_single_pair_iuv(dt, gt, ipoints, upoints, vpoints)
+            elif "embedding" in record:
+                return self.computeOgps_single_pair_cse(
+                    dt,
+                    gt,
+                    py,
+                    px,
+                    pt_mask,
+                    record["coarse_segm"],
+                    record["embedding"],
+                    record["bbox"],
+                )
+            else:
+                raise Exception(f"Unknown record format: {record}")
+        elif "embedding" in dt:
+            return self.computeOgps_single_pair_cse(
+                dt, gt, py, px, pt_mask, dt["coarse_segm"], dt["embedding"], dt["bbox"]
+            )
+        raise Exception(f"Unknown detection format: {dt}")
+
+    def extract_iuv_from_quantized(self, dt, gt, py, px, pt_mask):
+        densepose_results_quantized = dt["densepose"]
+        ipoints, upoints, vpoints = self._extract_iuv(
+            densepose_results_quantized.labels_uv_uint8.numpy(), py, px, gt
+        )
+        ipoints[pt_mask == -1] = 0
+        return ipoints, upoints, vpoints
+
+    def extract_iuv_from_raw(self, dt, gt, py, px, pt_mask):
+        labels_dt = resample_fine_and_coarse_segm_tensors_to_bbox(
+            dt["fine_segm"].unsqueeze(0),
+            dt["coarse_segm"].unsqueeze(0),
+            dt["bbox"],
+        )
+        uv = resample_uv_tensors_to_bbox(
+            dt["u"].unsqueeze(0), dt["v"].unsqueeze(0), labels_dt.squeeze(0), dt["bbox"]
+        )
+        labels_uv_uint8 = torch.cat((labels_dt.byte(), (uv * 255).clamp(0, 255).byte()))
+        ipoints, upoints, vpoints = self._extract_iuv(labels_uv_uint8.numpy(), py, px, gt)
+        ipoints[pt_mask == -1] = 0
+        return ipoints, upoints, vpoints
+
+    def computeOgps_single_pair_iuv(self, dt, gt, ipoints, upoints, vpoints):
+        cVertsGT, ClosestVertsGTTransformed = self.findAllClosestVertsGT(gt)
+        cVerts = self.findAllClosestVertsUV(upoints, vpoints, ipoints)
+        # Get pairwise geodesic distances between gt and estimated mesh points.
+        dist = self.getDistancesUV(ClosestVertsGTTransformed, cVerts)
+        # Compute the Ogps measure.
+        # Find the mean geodesic normalization distance for
+        # each GT point, based on which part it is on.
+        Current_Mean_Distances = self.Mean_Distances[
+            self.CoarseParts[self.Part_ids[cVertsGT[cVertsGT > 0].astype(int) - 1]]
+        ]
+        return dist, Current_Mean_Distances
+
+    def computeOgps_single_pair_cse(
+        self, dt, gt, py, px, pt_mask, coarse_segm, embedding, bbox_xywh_abs
+    ):
+        # 0-based mesh vertex indices
+        cVertsGT = torch.as_tensor(gt["dp_vertex"], dtype=torch.int64)
+        # label for each pixel of the bbox, [H, W] tensor of long
+        labels_dt = resample_coarse_segm_tensor_to_bbox(
+            coarse_segm.unsqueeze(0), bbox_xywh_abs
+        ).squeeze(0)
+        x, y, w, h = bbox_xywh_abs
+        # embedding for each pixel of the bbox, [D, H, W] tensor of float32
+        embedding = F.interpolate(
+            embedding.unsqueeze(0), (int(h), int(w)), mode="bilinear", align_corners=False
+        ).squeeze(0)
+        # valid locations py, px
+        py_pt = torch.from_numpy(py[pt_mask > -1])
+        px_pt = torch.from_numpy(px[pt_mask > -1])
+        cVerts = torch.ones_like(cVertsGT) * -1
+        cVerts[pt_mask > -1] = self.findClosestVertsCse(
+            embedding, py_pt, px_pt, labels_dt, gt["ref_model"]
+        )
+        # Get pairwise geodesic distances between gt and estimated mesh points.
+        dist = self.getDistancesCse(cVertsGT, cVerts, gt["ref_model"])
+        # normalize distances
+        if (gt["ref_model"] == "smpl_27554") and ("dp_I" in gt):
+            Current_Mean_Distances = self.Mean_Distances[
+                self.CoarseParts[np.array(gt["dp_I"], dtype=int)]
+            ]
+        else:
+            Current_Mean_Distances = 0.255
+        return dist, Current_Mean_Distances
+
+    def computeOgps(self, imgId, catId):
+        p = self.params
+        # dimension here should be Nxm
+        g = self._gts[imgId, catId]
+        d = self._dts[imgId, catId]
+        inds = np.argsort([-d_["score"] for d_ in d], kind="mergesort")
+        d = [d[i] for i in inds]
+        if len(d) > p.maxDets[-1]:
+            d = d[0 : p.maxDets[-1]]
+        # if len(gts) == 0 and len(dts) == 0:
+        if len(g) == 0 or len(d) == 0:
+            return []
+        ious = np.zeros((len(d), len(g)))
+        # compute opgs between each detection and ground truth object
+        # sigma = self.sigma #0.255 # dist = 0.3m corresponds to ogps = 0.5
+        # 1 # dist = 0.3m corresponds to ogps = 0.96
+        # 1.45 # dist = 1.7m (person height) corresponds to ogps = 0.5)
+        for j, gt in enumerate(g):
+            if not gt["ignore"]:
+                g_ = gt["bbox"]
+                for i, dt in enumerate(d):
+                    #
+                    dy = int(dt["bbox"][3])
+                    dx = int(dt["bbox"][2])
+                    dp_x = np.array(gt["dp_x"]) * g_[2] / 255.0
+                    dp_y = np.array(gt["dp_y"]) * g_[3] / 255.0
+                    py = (dp_y + g_[1] - dt["bbox"][1]).astype(int)
+                    px = (dp_x + g_[0] - dt["bbox"][0]).astype(int)
+                    #
+                    pts = np.zeros(len(px))
+                    pts[px >= dx] = -1
+                    pts[py >= dy] = -1
+                    pts[px < 0] = -1
+                    pts[py < 0] = -1
+                    if len(pts) < 1:
+                        ogps = 0.0
+                    elif np.max(pts) == -1:
+                        ogps = 0.0
+                    else:
+                        px[pts == -1] = 0
+                        py[pts == -1] = 0
+                        dists_between_matches, dist_norm_coeffs = self.computeOgps_single_pair(
+                            dt, gt, py, px, pts
+                        )
+                        # Compute gps
+                        ogps_values = np.exp(
+                            -(dists_between_matches**2) / (2 * (dist_norm_coeffs**2))
+                        )
+                        #
+                        ogps = np.mean(ogps_values) if len(ogps_values) > 0 else 0.0
+                    ious[i, j] = ogps
+
+        gbb = [gt["bbox"] for gt in g]
+        dbb = [dt["bbox"] for dt in d]
+
+        # compute iou between each dt and gt region
+        iscrowd = [int(o.get("iscrowd", 0)) for o in g]
+        ious_bb = maskUtils.iou(dbb, gbb, iscrowd)
+        return ious, ious_bb
+
+    def evaluateImg(self, imgId, catId, aRng, maxDet):
+        """
+        perform evaluation for single category and image
+        :return: dict (single image results)
+        """
+
+        p = self.params
+        if p.useCats:
+            gt = self._gts[imgId, catId]
+            dt = self._dts[imgId, catId]
+        else:
+            gt = [_ for cId in p.catIds for _ in self._gts[imgId, cId]]
+            dt = [_ for cId in p.catIds for _ in self._dts[imgId, cId]]
+        if len(gt) == 0 and len(dt) == 0:
+            return None
+
+        for g in gt:
+            # g['_ignore'] = g['ignore']
+            if g["ignore"] or (g["area"] < aRng[0] or g["area"] > aRng[1]):
+                g["_ignore"] = True
+            else:
+                g["_ignore"] = False
+
+        # sort dt highest score first, sort gt ignore last
+        gtind = np.argsort([g["_ignore"] for g in gt], kind="mergesort")
+        gt = [gt[i] for i in gtind]
+        dtind = np.argsort([-d["score"] for d in dt], kind="mergesort")
+        dt = [dt[i] for i in dtind[0:maxDet]]
+        iscrowd = [int(o.get("iscrowd", 0)) for o in gt]
+        # load computed ious
+        if p.iouType == "densepose":
+            # print('Checking the length', len(self.ious[imgId, catId]))
+            # if len(self.ious[imgId, catId]) == 0:
+            #    print(self.ious[imgId, catId])
+            ious = (
+                self.ious[imgId, catId][0][:, gtind]
+                if len(self.ious[imgId, catId]) > 0
+                else self.ious[imgId, catId]
+            )
+            ioubs = (
+                self.ious[imgId, catId][1][:, gtind]
+                if len(self.ious[imgId, catId]) > 0
+                else self.ious[imgId, catId]
+            )
+            if self._dpEvalMode in {DensePoseEvalMode.GPSM, DensePoseEvalMode.IOU}:
+                iousM = (
+                    self.real_ious[imgId, catId][:, gtind]
+                    if len(self.real_ious[imgId, catId]) > 0
+                    else self.real_ious[imgId, catId]
+                )
+        else:
+            ious = (
+                self.ious[imgId, catId][:, gtind]
+                if len(self.ious[imgId, catId]) > 0
+                else self.ious[imgId, catId]
+            )
+
+        T = len(p.iouThrs)
+        G = len(gt)
+        D = len(dt)
+        gtm = np.zeros((T, G))
+        dtm = np.zeros((T, D))
+        gtIg = np.array([g["_ignore"] for g in gt])
+        dtIg = np.zeros((T, D))
+        if np.all(gtIg) and p.iouType == "densepose":
+            dtIg = np.logical_or(dtIg, True)
+
+        if len(ious) > 0:  # and not p.iouType == 'densepose':
+            for tind, t in enumerate(p.iouThrs):
+                for dind, d in enumerate(dt):
+                    # information about best match so far (m=-1 -> unmatched)
+                    iou = min([t, 1 - 1e-10])
+                    m = -1
+                    for gind, _g in enumerate(gt):
+                        # if this gt already matched, and not a crowd, continue
+                        if gtm[tind, gind] > 0 and not iscrowd[gind]:
+                            continue
+                        # if dt matched to reg gt, and on ignore gt, stop
+                        if m > -1 and gtIg[m] == 0 and gtIg[gind] == 1:
+                            break
+                        if p.iouType == "densepose":
+                            if self._dpEvalMode == DensePoseEvalMode.GPSM:
+                                new_iou = np.sqrt(iousM[dind, gind] * ious[dind, gind])
+                            elif self._dpEvalMode == DensePoseEvalMode.IOU:
+                                new_iou = iousM[dind, gind]
+                            elif self._dpEvalMode == DensePoseEvalMode.GPS:
+                                new_iou = ious[dind, gind]
+                        else:
+                            new_iou = ious[dind, gind]
+                        if new_iou < iou:
+                            continue
+                        if new_iou == 0.0:
+                            continue
+                        # if match successful and best so far, store appropriately
+                        iou = new_iou
+                        m = gind
+                    # if match made store id of match for both dt and gt
+                    if m == -1:
+                        continue
+                    dtIg[tind, dind] = gtIg[m]
+                    dtm[tind, dind] = gt[m]["id"]
+                    gtm[tind, m] = d["id"]
+
+        if p.iouType == "densepose":
+            if not len(ioubs) == 0:
+                for dind, d in enumerate(dt):
+                    # information about best match so far (m=-1 -> unmatched)
+                    if dtm[tind, dind] == 0:
+                        ioub = 0.8
+                        m = -1
+                        for gind, _g in enumerate(gt):
+                            # if this gt already matched, and not a crowd, continue
+                            if gtm[tind, gind] > 0 and not iscrowd[gind]:
+                                continue
+                            # continue to next gt unless better match made
+                            if ioubs[dind, gind] < ioub:
+                                continue
+                            # if match successful and best so far, store appropriately
+                            ioub = ioubs[dind, gind]
+                            m = gind
+                            # if match made store id of match for both dt and gt
+                        if m > -1:
+                            dtIg[:, dind] = gtIg[m]
+                            if gtIg[m]:
+                                dtm[tind, dind] = gt[m]["id"]
+                                gtm[tind, m] = d["id"]
+        # set unmatched detections outside of area range to ignore
+        a = np.array([d["area"] < aRng[0] or d["area"] > aRng[1] for d in dt]).reshape((1, len(dt)))
+        dtIg = np.logical_or(dtIg, np.logical_and(dtm == 0, np.repeat(a, T, 0)))
+        # store results for given image and category
+        # print('Done with the function', len(self.ious[imgId, catId]))
+        return {
+            "image_id": imgId,
+            "category_id": catId,
+            "aRng": aRng,
+            "maxDet": maxDet,
+            "dtIds": [d["id"] for d in dt],
+            "gtIds": [g["id"] for g in gt],
+            "dtMatches": dtm,
+            "gtMatches": gtm,
+            "dtScores": [d["score"] for d in dt],
+            "gtIgnore": gtIg,
+            "dtIgnore": dtIg,
+        }
+
+    def accumulate(self, p=None):
+        """
+        Accumulate per image evaluation results and store the result in self.eval
+        :param p: input params for evaluation
+        :return: None
+        """
+        logger.info("Accumulating evaluation results...")
+        tic = time.time()
+        if not self.evalImgs:
+            logger.info("Please run evaluate() first")
+        # allows input customized parameters
+        if p is None:
+            p = self.params
+        p.catIds = p.catIds if p.useCats == 1 else [-1]
+        T = len(p.iouThrs)
+        R = len(p.recThrs)
+        K = len(p.catIds) if p.useCats else 1
+        A = len(p.areaRng)
+        M = len(p.maxDets)
+        precision = -(np.ones((T, R, K, A, M)))  # -1 for the precision of absent categories
+        recall = -(np.ones((T, K, A, M)))
+
+        # create dictionary for future indexing
+        logger.info("Categories: {}".format(p.catIds))
+        _pe = self._paramsEval
+        catIds = _pe.catIds if _pe.useCats else [-1]
+        setK = set(catIds)
+        setA = set(map(tuple, _pe.areaRng))
+        setM = set(_pe.maxDets)
+        setI = set(_pe.imgIds)
+        # get inds to evaluate
+        k_list = [n for n, k in enumerate(p.catIds) if k in setK]
+        m_list = [m for n, m in enumerate(p.maxDets) if m in setM]
+        a_list = [n for n, a in enumerate(map(lambda x: tuple(x), p.areaRng)) if a in setA]
+        i_list = [n for n, i in enumerate(p.imgIds) if i in setI]
+        I0 = len(_pe.imgIds)
+        A0 = len(_pe.areaRng)
+        # retrieve E at each category, area range, and max number of detections
+        for k, k0 in enumerate(k_list):
+            Nk = k0 * A0 * I0
+            for a, a0 in enumerate(a_list):
+                Na = a0 * I0
+                for m, maxDet in enumerate(m_list):
+                    E = [self.evalImgs[Nk + Na + i] for i in i_list]
+                    E = [e for e in E if e is not None]
+                    if len(E) == 0:
+                        continue
+                    dtScores = np.concatenate([e["dtScores"][0:maxDet] for e in E])
+
+                    # different sorting method generates slightly different results.
+                    # mergesort is used to be consistent as Matlab implementation.
+                    inds = np.argsort(-dtScores, kind="mergesort")
+
+                    dtm = np.concatenate([e["dtMatches"][:, 0:maxDet] for e in E], axis=1)[:, inds]
+                    dtIg = np.concatenate([e["dtIgnore"][:, 0:maxDet] for e in E], axis=1)[:, inds]
+                    gtIg = np.concatenate([e["gtIgnore"] for e in E])
+                    npig = np.count_nonzero(gtIg == 0)
+                    if npig == 0:
+                        continue
+                    tps = np.logical_and(dtm, np.logical_not(dtIg))
+                    fps = np.logical_and(np.logical_not(dtm), np.logical_not(dtIg))
+                    tp_sum = np.cumsum(tps, axis=1).astype(dtype=float)
+                    fp_sum = np.cumsum(fps, axis=1).astype(dtype=float)
+                    for t, (tp, fp) in enumerate(zip(tp_sum, fp_sum)):
+                        tp = np.array(tp)
+                        fp = np.array(fp)
+                        nd = len(tp)
+                        rc = tp / npig
+                        pr = tp / (fp + tp + np.spacing(1))
+                        q = np.zeros((R,))
+
+                        if nd:
+                            recall[t, k, a, m] = rc[-1]
+                        else:
+                            recall[t, k, a, m] = 0
+
+                        # numpy is slow without cython optimization for accessing elements
+                        # use python array gets significant speed improvement
+                        pr = pr.tolist()
+                        q = q.tolist()
+
+                        for i in range(nd - 1, 0, -1):
+                            if pr[i] > pr[i - 1]:
+                                pr[i - 1] = pr[i]
+
+                        inds = np.searchsorted(rc, p.recThrs, side="left")
+                        try:
+                            for ri, pi in enumerate(inds):
+                                q[ri] = pr[pi]
+                        except Exception:
+                            pass
+                        precision[t, :, k, a, m] = np.array(q)
+        logger.info(
+            "Final: max precision {}, min precision {}".format(np.max(precision), np.min(precision))
+        )
+        self.eval = {
+            "params": p,
+            "counts": [T, R, K, A, M],
+            "date": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+            "precision": precision,
+            "recall": recall,
+        }
+        toc = time.time()
+        logger.info("DONE (t={:0.2f}s).".format(toc - tic))
+
+    def summarize(self):
+        """
+        Compute and display summary metrics for evaluation results.
+        Note this function can *only* be applied on the default parameter setting
+        """
+
+        def _summarize(ap=1, iouThr=None, areaRng="all", maxDets=100):
+            p = self.params
+            iStr = " {:<18} {} @[ {}={:<9} | area={:>6s} | maxDets={:>3d} ] = {:0.3f}"
+            titleStr = "Average Precision" if ap == 1 else "Average Recall"
+            typeStr = "(AP)" if ap == 1 else "(AR)"
+            measure = "IoU"
+            if self.params.iouType == "keypoints":
+                measure = "OKS"
+            elif self.params.iouType == "densepose":
+                measure = "OGPS"
+            iouStr = (
+                "{:0.2f}:{:0.2f}".format(p.iouThrs[0], p.iouThrs[-1])
+                if iouThr is None
+                else "{:0.2f}".format(iouThr)
+            )
+
+            aind = [i for i, aRng in enumerate(p.areaRngLbl) if aRng == areaRng]
+            mind = [i for i, mDet in enumerate(p.maxDets) if mDet == maxDets]
+            if ap == 1:
+                # dimension of precision: [TxRxKxAxM]
+                s = self.eval["precision"]
+                # IoU
+                if iouThr is not None:
+                    t = np.where(np.abs(iouThr - p.iouThrs) < 0.001)[0]
+                    s = s[t]
+                s = s[:, :, :, aind, mind]
+            else:
+                # dimension of recall: [TxKxAxM]
+                s = self.eval["recall"]
+                if iouThr is not None:
+                    t = np.where(np.abs(iouThr - p.iouThrs) < 0.001)[0]
+                    s = s[t]
+                s = s[:, :, aind, mind]
+            if len(s[s > -1]) == 0:
+                mean_s = -1
+            else:
+                mean_s = np.mean(s[s > -1])
+            logger.info(iStr.format(titleStr, typeStr, measure, iouStr, areaRng, maxDets, mean_s))
+            return mean_s
+
+        def _summarizeDets():
+            stats = np.zeros((12,))
+            stats[0] = _summarize(1)
+            stats[1] = _summarize(1, iouThr=0.5, maxDets=self.params.maxDets[2])
+            stats[2] = _summarize(1, iouThr=0.75, maxDets=self.params.maxDets[2])
+            stats[3] = _summarize(1, areaRng="small", maxDets=self.params.maxDets[2])
+            stats[4] = _summarize(1, areaRng="medium", maxDets=self.params.maxDets[2])
+            stats[5] = _summarize(1, areaRng="large", maxDets=self.params.maxDets[2])
+            stats[6] = _summarize(0, maxDets=self.params.maxDets[0])
+            stats[7] = _summarize(0, maxDets=self.params.maxDets[1])
+            stats[8] = _summarize(0, maxDets=self.params.maxDets[2])
+            stats[9] = _summarize(0, areaRng="small", maxDets=self.params.maxDets[2])
+            stats[10] = _summarize(0, areaRng="medium", maxDets=self.params.maxDets[2])
+            stats[11] = _summarize(0, areaRng="large", maxDets=self.params.maxDets[2])
+            return stats
+
+        def _summarizeKps():
+            stats = np.zeros((10,))
+            stats[0] = _summarize(1, maxDets=20)
+            stats[1] = _summarize(1, maxDets=20, iouThr=0.5)
+            stats[2] = _summarize(1, maxDets=20, iouThr=0.75)
+            stats[3] = _summarize(1, maxDets=20, areaRng="medium")
+            stats[4] = _summarize(1, maxDets=20, areaRng="large")
+            stats[5] = _summarize(0, maxDets=20)
+            stats[6] = _summarize(0, maxDets=20, iouThr=0.5)
+            stats[7] = _summarize(0, maxDets=20, iouThr=0.75)
+            stats[8] = _summarize(0, maxDets=20, areaRng="medium")
+            stats[9] = _summarize(0, maxDets=20, areaRng="large")
+            return stats
+
+        def _summarizeUvs():
+            stats = [_summarize(1, maxDets=self.params.maxDets[0])]
+            min_threshold = self.params.iouThrs.min()
+            if min_threshold <= 0.201:
+                stats += [_summarize(1, maxDets=self.params.maxDets[0], iouThr=0.2)]
+            if min_threshold <= 0.301:
+                stats += [_summarize(1, maxDets=self.params.maxDets[0], iouThr=0.3)]
+            if min_threshold <= 0.401:
+                stats += [_summarize(1, maxDets=self.params.maxDets[0], iouThr=0.4)]
+            stats += [
+                _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.5),
+                _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.75),
+                _summarize(1, maxDets=self.params.maxDets[0], areaRng="medium"),
+                _summarize(1, maxDets=self.params.maxDets[0], areaRng="large"),
+                _summarize(0, maxDets=self.params.maxDets[0]),
+                _summarize(0, maxDets=self.params.maxDets[0], iouThr=0.5),
+                _summarize(0, maxDets=self.params.maxDets[0], iouThr=0.75),
+                _summarize(0, maxDets=self.params.maxDets[0], areaRng="medium"),
+                _summarize(0, maxDets=self.params.maxDets[0], areaRng="large"),
+            ]
+            return np.array(stats)
+
+        def _summarizeUvsOld():
+            stats = np.zeros((18,))
+            stats[0] = _summarize(1, maxDets=self.params.maxDets[0])
+            stats[1] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.5)
+            stats[2] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.55)
+            stats[3] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.60)
+            stats[4] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.65)
+            stats[5] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.70)
+            stats[6] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.75)
+            stats[7] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.80)
+            stats[8] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.85)
+            stats[9] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.90)
+            stats[10] = _summarize(1, maxDets=self.params.maxDets[0], iouThr=0.95)
+            stats[11] = _summarize(1, maxDets=self.params.maxDets[0], areaRng="medium")
+            stats[12] = _summarize(1, maxDets=self.params.maxDets[0], areaRng="large")
+            stats[13] = _summarize(0, maxDets=self.params.maxDets[0])
+            stats[14] = _summarize(0, maxDets=self.params.maxDets[0], iouThr=0.5)
+            stats[15] = _summarize(0, maxDets=self.params.maxDets[0], iouThr=0.75)
+            stats[16] = _summarize(0, maxDets=self.params.maxDets[0], areaRng="medium")
+            stats[17] = _summarize(0, maxDets=self.params.maxDets[0], areaRng="large")
+            return stats
+
+        if not self.eval:
+            raise Exception("Please run accumulate() first")
+        iouType = self.params.iouType
+        if iouType in ["segm", "bbox"]:
+            summarize = _summarizeDets
+        elif iouType in ["keypoints"]:
+            summarize = _summarizeKps
+        elif iouType in ["densepose"]:
+            summarize = _summarizeUvs
+        self.stats = summarize()
+
+    def __str__(self):
+        self.summarize()
+
+    # ================ functions for dense pose ==============================
+    def findAllClosestVertsUV(self, U_points, V_points, Index_points):
+        ClosestVerts = np.ones(Index_points.shape) * -1
+        for i in np.arange(24):
+            #
+            if (i + 1) in Index_points:
+                UVs = np.array(
+                    [U_points[Index_points == (i + 1)], V_points[Index_points == (i + 1)]]
+                )
+                Current_Part_UVs = self.Part_UVs[i]
+                Current_Part_ClosestVertInds = self.Part_ClosestVertInds[i]
+                D = ssd.cdist(Current_Part_UVs.transpose(), UVs.transpose()).squeeze()
+                ClosestVerts[Index_points == (i + 1)] = Current_Part_ClosestVertInds[
+                    np.argmin(D, axis=0)
+                ]
+        ClosestVertsTransformed = self.PDIST_transform[ClosestVerts.astype(int) - 1]
+        ClosestVertsTransformed[ClosestVerts < 0] = 0
+        return ClosestVertsTransformed
+
+    def findClosestVertsCse(self, embedding, py, px, mask, mesh_name):
+        mesh_vertex_embeddings = self.embedder(mesh_name)
+        pixel_embeddings = embedding[:, py, px].t().to(device="cuda")
+        mask_vals = mask[py, px]
+        edm = squared_euclidean_distance_matrix(pixel_embeddings, mesh_vertex_embeddings)
+        vertex_indices = edm.argmin(dim=1).cpu()
+        vertex_indices[mask_vals <= 0] = -1
+        return vertex_indices
+
+    def findAllClosestVertsGT(self, gt):
+        #
+        I_gt = np.array(gt["dp_I"])
+        U_gt = np.array(gt["dp_U"])
+        V_gt = np.array(gt["dp_V"])
+        #
+        # print(I_gt)
+        #
+        ClosestVertsGT = np.ones(I_gt.shape) * -1
+        for i in np.arange(24):
+            if (i + 1) in I_gt:
+                UVs = np.array([U_gt[I_gt == (i + 1)], V_gt[I_gt == (i + 1)]])
+                Current_Part_UVs = self.Part_UVs[i]
+                Current_Part_ClosestVertInds = self.Part_ClosestVertInds[i]
+                D = ssd.cdist(Current_Part_UVs.transpose(), UVs.transpose()).squeeze()
+                ClosestVertsGT[I_gt == (i + 1)] = Current_Part_ClosestVertInds[np.argmin(D, axis=0)]
+        #
+        ClosestVertsGTTransformed = self.PDIST_transform[ClosestVertsGT.astype(int) - 1]
+        ClosestVertsGTTransformed[ClosestVertsGT < 0] = 0
+        return ClosestVertsGT, ClosestVertsGTTransformed
+
+    def getDistancesCse(self, cVertsGT, cVerts, mesh_name):
+        geodists_vertices = torch.ones_like(cVertsGT) * float("inf")
+        selected = (cVertsGT >= 0) * (cVerts >= 0)
+        mesh = create_mesh(mesh_name, "cpu")
+        geodists_vertices[selected] = mesh.geodists[cVertsGT[selected], cVerts[selected]]
+        return geodists_vertices.numpy()
+
+    def getDistancesUV(self, cVertsGT, cVerts):
+        #
+        n = 27554
+        dists = []
+        for d in range(len(cVertsGT)):
+            if cVertsGT[d] > 0:
+                if cVerts[d] > 0:
+                    i = cVertsGT[d] - 1
+                    j = cVerts[d] - 1
+                    if j == i:
+                        dists.append(0)
+                    elif j > i:
+                        ccc = i
+                        i = j
+                        j = ccc
+                        i = n - i - 1
+                        j = n - j - 1
+                        k = (n * (n - 1) / 2) - (n - i) * ((n - i) - 1) / 2 + j - i - 1
+                        k = (n * n - n) / 2 - k - 1
+                        dists.append(self.Pdist_matrix[int(k)][0])
+                    else:
+                        i = n - i - 1
+                        j = n - j - 1
+                        k = (n * (n - 1) / 2) - (n - i) * ((n - i) - 1) / 2 + j - i - 1
+                        k = (n * n - n) / 2 - k - 1
+                        dists.append(self.Pdist_matrix[int(k)][0])
+                else:
+                    dists.append(np.inf)
+        return np.atleast_1d(np.array(dists).squeeze())
+
+
+class Params:
+    """
+    Params for coco evaluation api
+    """
+
+    def setDetParams(self):
+        self.imgIds = []
+        self.catIds = []
+        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
+        self.iouThrs = np.linspace(0.5, 0.95, int(np.round((0.95 - 0.5) / 0.05)) + 1, endpoint=True)
+        self.recThrs = np.linspace(0.0, 1.00, int(np.round((1.00 - 0.0) / 0.01)) + 1, endpoint=True)
+        self.maxDets = [1, 10, 100]
+        self.areaRng = [
+            [0**2, 1e5**2],
+            [0**2, 32**2],
+            [32**2, 96**2],
+            [96**2, 1e5**2],
+        ]
+        self.areaRngLbl = ["all", "small", "medium", "large"]
+        self.useCats = 1
+
+    def setKpParams(self):
+        self.imgIds = []
+        self.catIds = []
+        # np.arange causes trouble.  the data point on arange is slightly larger than the true value
+        self.iouThrs = np.linspace(0.5, 0.95, np.round((0.95 - 0.5) / 0.05) + 1, endpoint=True)
+        self.recThrs = np.linspace(0.0, 1.00, np.round((1.00 - 0.0) / 0.01) + 1, endpoint=True)
+        self.maxDets = [20]
+        self.areaRng = [[0**2, 1e5**2], [32**2, 96**2], [96**2, 1e5**2]]
+        self.areaRngLbl = ["all", "medium", "large"]
+        self.useCats = 1
+
+    def setUvParams(self):
+        self.imgIds = []
+        self.catIds = []
+        self.iouThrs = np.linspace(0.5, 0.95, int(np.round((0.95 - 0.5) / 0.05)) + 1, endpoint=True)
+        self.recThrs = np.linspace(0.0, 1.00, int(np.round((1.00 - 0.0) / 0.01)) + 1, endpoint=True)
+        self.maxDets = [20]
+        self.areaRng = [[0**2, 1e5**2], [32**2, 96**2], [96**2, 1e5**2]]
+        self.areaRngLbl = ["all", "medium", "large"]
+        self.useCats = 1
+
+    def __init__(self, iouType="segm"):
+        if iouType == "segm" or iouType == "bbox":
+            self.setDetParams()
+        elif iouType == "keypoints":
+            self.setKpParams()
+        elif iouType == "densepose":
+            self.setUvParams()
+        else:
+            raise Exception("iouType not supported")
+        self.iouType = iouType
+        # useSegm is deprecated
+        self.useSegm = None