Fix CXCYWH to XYXY conversion for integer bounding boxes (#9322)

Co-authored-by: zy1git <zycoding1@gmail.com>
Co-authored-by: Zhitao Yu <zhitao@fb.com>

Author: 3 people

test/test_transforms_v2.py

@@ -4412,6 +4412,27 @@ def test_errors(self):
                 input_tv_tensor, old_format=input_tv_tensor.format, new_format=input_tv_tensor.format
             )
 
+    def test_cxcywh_to_xyxy_odd_dimensions(self):
+        # Non-regression test for https://github.com/pytorch/vision/issues/8887
+        # Integer bounding boxes with odd width/height produced incorrect results
+        # due to integer division rounding issues (ceil instead of truncation).
+        bounding_boxes = tv_tensors.BoundingBoxes(
+            [[5, 6, 10, 13]],
+            format=tv_tensors.BoundingBoxFormat.CXCYWH,
+            canvas_size=(17, 11),
+            dtype=torch.int64,
+        )
+
+        actual = F.convert_bounding_box_format(bounding_boxes, new_format=tv_tensors.BoundingBoxFormat.XYXY)
+        expected = tv_tensors.BoundingBoxes(
+            [[0, 0, 10, 12]],
+            format=tv_tensors.BoundingBoxFormat.XYXY,
+            canvas_size=(17, 11),
+        )
+
+        assert (actual >= 0).all()
+        torch.testing.assert_close(actual, expected)
+
 
 class TestResizedCrop:
     INPUT_SIZE = (17, 11)

torchvision/transforms/v2/functional/_meta.py

@@ -177,17 +177,29 @@ def _xyxy_to_xywh(xyxy: torch.Tensor, inplace: bool) -> torch.Tensor:
 
 
 def _cxcywh_to_xyxy(cxcywh: torch.Tensor, inplace: bool) -> torch.Tensor:
-    if not inplace:
+    # For integer tensors, use float arithmetic to match the behavior of
+    # `torchvision.ops._box_convert._box_cxcywh_to_xyxy`.
+    original = cxcywh
+    dtype = cxcywh.dtype
+    need_cast = not cxcywh.is_floating_point()
+
+    if need_cast:
+        cxcywh = cxcywh.float()
+    elif not inplace:
         cxcywh = cxcywh.clone()
 
-    # Trick to do fast division by 2 and ceil, without casting. It produces the same result as
-    # `torchvision.ops._box_convert._box_cxcywh_to_xyxy`.
-    half_wh = cxcywh[..., 2:].div(-2, rounding_mode=None if cxcywh.is_floating_point() else "floor").abs_()
+    half_wh = cxcywh[..., 2:] / 2
     # (cx - width / 2) = x1, same for y1
     cxcywh[..., :2].sub_(half_wh)
     # (x1 + width) = x2, same for y2
     cxcywh[..., 2:].add_(cxcywh[..., :2])
 
+    if need_cast:
+        cxcywh = cxcywh.to(dtype)
+        if inplace:
+            original[:] = cxcywh
+            return original
+
     return cxcywh