[fbsync] Added Middlebury2014 dataset (#6350)

Summary:
* Added Middlebury2014   dataset

* adressed nits

* mypy fix

Reviewed By: YosuaMichael

Differential Revision: D39426991

fbshipit-source-id: 382a30aee1b4927a5cb47a338f2307ec3241bb71

Author: datumbox

docs/source/datasets.rst

@@ -117,6 +117,7 @@ Stereo Matching
     SintelStereo
     InStereo2k
     ETH3DStereo
+    Middlebury2014Stereo
 
 Image pairs
 ~~~~~~~~~~~

test/test_datasets.py

@@ -3218,5 +3218,98 @@ def test_bad_input(self):
                 pass
 
 
+class Middlebury2014StereoTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.Middlebury2014Stereo
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(
+        split=("train", "additional"),
+        calibration=("perfect", "imperfect", "both"),
+        use_ambient_views=(True, False),
+    )
+    FEATURE_TYPES = (PIL.Image.Image, PIL.Image.Image, (np.ndarray, type(None)), (np.ndarray, type(None)))
+
+    @staticmethod
+    def _make_scene_folder(root_dir: str, scene_name: str, split: str) -> None:
+        calibrations = [None] if split == "test" else ["-perfect", "-imperfect"]
+        root_dir = pathlib.Path(root_dir)
+
+        for c in calibrations:
+            scene_dir = root_dir / f"{scene_name}{c}"
+            os.makedirs(scene_dir, exist_ok=True)
+            # make normal images first
+            datasets_utils.create_image_file(root=scene_dir, name="im0.png", size=(3, 100, 100))
+            datasets_utils.create_image_file(root=scene_dir, name="im1.png", size=(3, 100, 100))
+            datasets_utils.create_image_file(root=scene_dir, name="im1E.png", size=(3, 100, 100))
+            datasets_utils.create_image_file(root=scene_dir, name="im1L.png", size=(3, 100, 100))
+            # these are going to end up being gray scale images
+            datasets_utils.make_fake_pfm_file(h=100, w=100, file_name=scene_dir / "disp0.pfm")
+            datasets_utils.make_fake_pfm_file(h=100, w=100, file_name=scene_dir / "disp1.pfm")
+
+    def inject_fake_data(self, tmpdir, config):
+        split_scene_map = {
+            "train": ["Adirondack", "Jadeplant", "Motorcycle", "Piano"],
+            "additional": ["Backpack", "Bicycle1", "Cable", "Classroom1"],
+            "test": ["Plants", "Classroom2E", "Classroom2", "Australia"],
+        }
+
+        middlebury_dir = pathlib.Path(tmpdir, "Middlebury2014")
+        os.makedirs(middlebury_dir, exist_ok=True)
+
+        split_dir = middlebury_dir / config["split"]
+        os.makedirs(split_dir, exist_ok=True)
+
+        num_examples = {"train": 2, "additional": 3, "test": 4}.get(config["split"], 0)
+        for idx in range(num_examples):
+            scene_name = split_scene_map[config["split"]][idx]
+            self._make_scene_folder(root_dir=split_dir, scene_name=scene_name, split=config["split"])
+
+        if config["calibration"] == "both":
+            num_examples *= 2
+        return num_examples
+
+    def test_train_splits(self):
+        for split, calibration in itertools.product(["train", "additional"], ["perfect", "imperfect", "both"]):
+            with self.create_dataset(split=split, calibration=calibration) as (dataset, _):
+                for left, right, disparity, mask in dataset:
+                    datasets_utils.shape_test_for_stereo(left, right, disparity, mask)
+
+    def test_test_split(self):
+        for split in ["test"]:
+            with self.create_dataset(split=split, calibration=None) as (dataset, _):
+                for left, right, disparity, mask in dataset:
+                    datasets_utils.shape_test_for_stereo(left, right)
+
+    def test_augmented_view_usage(self):
+        with self.create_dataset(split="train", use_ambient_views=True) as (dataset, _):
+            for left, right, disparity, mask in dataset:
+                datasets_utils.shape_test_for_stereo(left, right, disparity, mask)
+
+    def test_value_err_train(self):
+        # train set invalid
+        split = "train"
+        calibration = None
+        with pytest.raises(
+            ValueError,
+            match=f"Split '{split}' has calibration settings, however None was provided as an argument."
+            f"\nSetting calibration to 'perfect' for split '{split}'. Available calibration settings are: 'perfect', 'imperfect', 'both'.",
+        ):
+            with self.create_dataset(split=split, calibration=calibration):
+                pass
+
+    def test_value_err_test(self):
+        # test set invalid
+        split = "test"
+        calibration = "perfect"
+        with pytest.raises(
+            ValueError, match="Split 'test' has only no calibration settings, please set `calibration=None`."
+        ):
+            with self.create_dataset(split=split, calibration=calibration):
+                pass
+
+    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

@@ -7,6 +7,7 @@
     InStereo2k,
     Kitti2012Stereo,
     Kitti2015Stereo,
+    Middlebury2014Stereo,
     SceneFlowStereo,
     SintelStereo,
 )
@@ -119,6 +120,7 @@
     "Kitti2012Stereo",
     "Kitti2015Stereo",
     "CarlaStereo",
+    "Middlebury2014Stereo",
     "CREStereo",
     "FallingThingsStereo",
     "SceneFlowStereo",