densepose/data/dataset_mapper.py

# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.

# pyre-unsafe

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