Added ETH3D stereo dataset

docs/source/datasets.rst

@@ -111,6 +111,7 @@ Stereo Matching
     CarlaStereo
     Kitti2012Stereo
     Kitti2015Stereo
+    ETH3DStereo
 
 Image pairs
 ~~~~~~~~~~~

test/test_datasets.py

@@ -2841,5 +2841,75 @@ def test_train_splits(self):
                 datasets_utils.shape_test_for_stereo(left, right, disparity)
 
 
+class ETH3DStereoestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.ETH3DStereo
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(split=("train", "test"))
+    FEATURE_TYPES = (PIL.Image.Image, PIL.Image.Image, (np.ndarray, type(None)), (np.ndarray, type(None)))
+
+    @staticmethod
+    def _create_scene_folder(num_examples: int, root_dir: str):
+        # make the root_dir if it does not exits
+        root_dir = pathlib.Path(root_dir)
+        os.makedirs(root_dir, exist_ok=True)
+
+        for i in range(num_examples):
+            scene_dir = root_dir / f"scene_{i}"
+            os.makedirs(scene_dir, exist_ok=True)
+            # populate with left right images
+            datasets_utils.create_image_file(root=scene_dir, name="im0.png", size=(100, 100))
+            datasets_utils.create_image_file(root=scene_dir, name="im1.png", size=(100, 100))
+
+    @staticmethod
+    def _create_annotation_folder(num_examples: int, root_dir: str):
+        # make the root_dir if it does not exits
+        root_dir = pathlib.Path(root_dir)
+        os.makedirs(root_dir, exist_ok=True)
+
+        # create scene directories
+        for i in range(num_examples):
+            scene_dir = root_dir / f"scene_{i}"
+            os.makedirs(scene_dir, exist_ok=True)
+            # populate with a random png file for occlusion mask, and a pfm file for disparity
+            datasets_utils.create_image_file(root=scene_dir, name="mask0nocc.png", size=(1, 100, 100))
+
+            pfm_path = scene_dir / "disp0GT.pfm"
+            datasets_utils.make_fake_pfm_file(h=100, w=100, file_name=pfm_path)
+
+    def inject_fake_data(self, tmpdir, config):
+        eth3d_dir = pathlib.Path(tmpdir) / "ETH3D"
+
+        num_examples = 2 if config["split"] == "train" else 3
+
+        split_name = "two_view_training" if config["split"] == "train" else "two_view_test"
+        split_dir = eth3d_dir / split_name
+        self._create_scene_folder(num_examples, split_dir)
+
+        if config["split"] == "train":
+            annot_dir = eth3d_dir / "two_view_training_gt"
+            self._create_annotation_folder(num_examples, annot_dir)
+
+        return num_examples
+
+    def test_training_splits(self):
+        with self.create_dataset(split="train") as (dataset, _):
+            assert dataset._images and len(dataset._images) == len(
+                dataset._disparities
+            ), "Training images do not match with training disparities"
+            for left, right, disparity, valid_mask in dataset:
+                datasets_utils.shape_test_for_stereo(left, right, disparity, valid_mask)
+
+    def test_testing_splits(self):
+        with self.create_dataset(split="test") as (dataset, _):
+            assert all(d == (None, None) for d in dataset._disparities)
+            for left, right, disparity, valid_mask in dataset:
+                assert valid_mask is None
+                datasets_utils.shape_test_for_stereo(left, right, disparity)
+
+    def test_bad_input(self):
+        with pytest.raises(ValueError, match="Unknown value 'bad' for argument split"):
+            with self.create_dataset(split="bad"):
+                pass
+
+
 if __name__ == "__main__":
     unittest.main()

torchvision/datasets/__init__.py

@@ -1,5 +1,5 @@
 from ._optical_flow import FlyingChairs, FlyingThings3D, HD1K, KittiFlow, Sintel
-from ._stereo_matching import CarlaStereo, Kitti2012Stereo, Kitti2015Stereo
+from ._stereo_matching import CarlaStereo, ETH3DStereo, Kitti2012Stereo, Kitti2015Stereo
 from .caltech import Caltech101, Caltech256
 from .celeba import CelebA
 from .cifar import CIFAR10, CIFAR100
@@ -109,4 +109,5 @@
     "Kitti2012Stereo",
     "Kitti2015Stereo",
     "CarlaStereo",
+    "ETH3DStereo",
 )

torchvision/datasets/_stereo_matching.py

@@ -359,3 +359,103 @@ def __getitem__(self, index: int) -> Tuple:
             Both ``disparity`` and ``valid_mask`` are ``None`` if the dataset split is test.
         """
         return super().__getitem__(index)
+
+
+class ETH3DStereo(StereoMatchingDataset):
+    """ETH3D `Low-Res Two-View <https://www.eth3d.net/datasets>`_ dataset.
+
+    The dataset is expected to have the following structure: ::
+
+        root
+            ETH3D
+                two_view_training
+                    scene1
+                        im1.png
+                        im0.png
+                        images.txt
+                        cameras.txt
+                        calib.txt
+                    scene2
+                        im1.png
+                        im0.png
+                        images.txt
+                        cameras.txt
+                        calib.txt
+                    ...
+                two_view_training_gt
+                    scene1
+                        disp0GT.pfm
+                        mask0nocc.png
+                    scene2
+                        disp0GT.pfm
+                        mask0nocc.png
+                    ...
+                two_view_testing
+                    scene1
+                        im1.png
+                        im0.png
+                        images.txt
+                        cameras.txt
+                        calib.txt
+                    scene2
+                        im1.png
+                        im0.png
+                        images.txt
+                        cameras.txt
+                        calib.txt
+                    ...
+
+    Args:
+        root (string): Root directory of the ETH3D Dataset.
+        split (string, optional): The dataset split of scenes, either "train" (default) or "test".
+        transforms (callable, optional): A function/transform that takes in a sample and returns a transformed version.
+    """
+
+    _has_built_in_disparity_mask = True
+
+    def __init__(self, root: str, split: str = "train", transforms: Optional[Callable] = None):
+        super().__init__(root, transforms)
+
+        verify_str_arg(split, "split", valid_values=("train", "test"))
+
+        root = Path(root) / "ETH3D"
+
+        img_dir = "two_view_training" if split == "train" else "two_view_test"
+        anot_dir = "two_view_training_gt"
+
+        left_img_pattern = str(root / img_dir / "*" / "im0.png")
+        right_img_pattern = str(root / img_dir / "*" / "im1.png")
+        self._images = self._scan_pairs(left_img_pattern, right_img_pattern)
+
+        if split == "test":
+            self._disparities = list((None, None) for _ in self._images)
+        else:
+            disparity_pattern = str(root / anot_dir / "*" / "disp0GT.pfm")
+            self._disparities = self._scan_pairs(disparity_pattern, None)
+
+    def _read_disparity(self, file_path: str) -> Tuple:
+        # test split has no disparity maps
+        if file_path is None:
+            return None, None
+
+        disparity_map = _read_pfm_file(file_path)
+        disparity_map = np.abs(disparity_map)  # ensure that the disparity is positive
+        mask_path = Path(file_path).parent / "mask0nocc.png"
+        valid_mask = Image.open(mask_path)
+        valid_mask = np.asarray(valid_mask).astype(np.bool_)
+        return disparity_map, valid_mask
+
+    def __getitem__(self, index: int) -> Tuple:
+        """Return example at given index.
+
+        Args:
+            index(int): The index of the example to retrieve
+
+        Returns:
+            tuple: A 4-tuple with ``(img_left, img_right, disparity, valid_mask)``.
+            The disparity is a numpy array of shape (1, H, W) and the images are PIL images.
+            ``valid_mask`` is implicitly ``None`` if the ``transforms`` parameter does not
+            generate a valid mask.
+            Both ``disparity`` and ``valid_mask`` are ``None`` if the dataset split is test.
+        """
+        return super().__getitem__(index)