[proto] Speed-up crop on bboxes and tests (#6881)

* [proto] Speed-up crop on bboxes and tests

* Fix linter

* Update _geometry.py

* Fixed device issue

* Revert changes in test/prototype_transforms_kernel_infos.py

* Fixed failing correctness tests

test/prototype_transforms_kernel_infos.py

@@ -862,6 +862,27 @@ def sample_inputs_crop_video():
         yield ArgsKwargs(video_loader, top=4, left=3, height=7, width=8)
 
 
+def reference_crop_bounding_box(bounding_box, *, format, top, left, height, width):
+
+    affine_matrix = np.array(
+        [
+            [1, 0, -left],
+            [0, 1, -top],
+        ],
+        dtype="float32",
+    )
+
+    expected_bboxes = reference_affine_bounding_box_helper(bounding_box, format=format, affine_matrix=affine_matrix)
+    return expected_bboxes, (height, width)
+
+
+def reference_inputs_crop_bounding_box():
+    for bounding_box_loader, params in itertools.product(
+        make_bounding_box_loaders(extra_dims=((), (4,))), [_CROP_PARAMS[0], _CROP_PARAMS[-1]]
+    ):
+        yield ArgsKwargs(bounding_box_loader, format=bounding_box_loader.format, **params)
+
+
 KERNEL_INFOS.extend(
     [
         KernelInfo(
@@ -875,6 +896,8 @@ KERNEL_INFOS.extend(
         KernelInfo(
             F.crop_bounding_box,
             sample_inputs_fn=sample_inputs_crop_bounding_box,
+            reference_fn=reference_crop_bounding_box,
+            reference_inputs_fn=reference_inputs_crop_bounding_box,
         ),
         KernelInfo(
             F.crop_mask,

test/test_prototype_transforms_functional.py

@@ -900,7 +900,8 @@ def test_correctness_center_crop_bounding_box(device, output_size):
     def _compute_expected_bbox(bbox, output_size_):
         format_ = bbox.format
         spatial_size_ = bbox.spatial_size
-        bbox = convert_format_bounding_box(bbox, format_, features.BoundingBoxFormat.XYWH)
+        dtype = bbox.dtype
+        bbox = convert_format_bounding_box(bbox.float(), format_, features.BoundingBoxFormat.XYWH)
 
         if len(output_size_) == 1:
             output_size_.append(output_size_[-1])
@@ -913,14 +914,9 @@ def test_correctness_center_crop_bounding_box(device, output_size):
             bbox[2].item(),
             bbox[3].item(),
         ]
-        out_bbox = features.BoundingBox(
-            out_bbox,
-            format=features.BoundingBoxFormat.XYWH,
-            spatial_size=output_size_,
-            dtype=bbox.dtype,
-            device=bbox.device,
-        )
-        return convert_format_bounding_box(out_bbox, features.BoundingBoxFormat.XYWH, format_)
+        out_bbox = torch.tensor(out_bbox)
+        out_bbox = convert_format_bounding_box(out_bbox, features.BoundingBoxFormat.XYWH, format_)
+        return out_bbox.to(dtype=dtype, device=bbox.device)
 
     for bboxes in make_bounding_boxes(extra_dims=((4,),)):
         bboxes = bboxes.to(device)

torchvision/prototype/transforms/functional/_geometry.py

@@ -802,22 +802,17 @@ def crop_bounding_box(
     height: int,
     width: int,
 ) -> Tuple[torch.Tensor, Tuple[int, int]]:
-    # TODO: Investigate if it makes sense from a performance perspective to have an implementation for every
-    #  BoundingBoxFormat instead of converting back and forth
-    bounding_box = convert_format_bounding_box(
-        bounding_box.clone(), old_format=format, new_format=features.BoundingBoxFormat.XYXY, inplace=True
-    )
+
+    bounding_box = bounding_box.clone()
 
     # Crop or implicit pad if left and/or top have negative values:
-    bounding_box[..., 0::2] -= left
-    bounding_box[..., 1::2] -= top
+    if format == features.BoundingBoxFormat.XYXY:
+        sub = torch.tensor([left, top, left, top], device=bounding_box.device)
+    else:
+        sub = torch.tensor([left, top, 0, 0], device=bounding_box.device)
+    bounding_box = bounding_box.sub_(sub)
 
-    return (
-        convert_format_bounding_box(
-            bounding_box, old_format=features.BoundingBoxFormat.XYXY, new_format=format, inplace=True
-        ),
-        (height, width),
-    )
+    return bounding_box, (height, width)
 
 
 def crop_mask(mask: torch.Tensor, top: int, left: int, height: int, width: int) -> torch.Tensor: