Skip to content

Fix int64 bbox rounding in perspective and elastic transforms #9354

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
raimbekovm wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix int64 bbox rounding in perspective and elastic transforms #9354

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0ffe6a6
Fix int64 bbox rounding in perspective and elastic transforms
raimbekovm Jan 25, 2026
d1d1604
Merge branch 'main' into fix/perspective-elastic-int64-bbox
zy1git Jan 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 15 additions & 9 deletions torchvision/transforms/v2/functional/_geometry.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2113,12 +2113,13 @@ def perspective_bounding_boxes(
original_dtype = bounding_boxes.dtype
is_rotated = tv_tensors.is_rotated_bounding_format(format)
intermediate_format = tv_tensors.BoundingBoxFormat.XYXYXYXY if is_rotated else tv_tensors.BoundingBoxFormat.XYXY
# TODO: first cast to float if bbox is int64 before convert_bounding_box_format
need_cast = not bounding_boxes.is_floating_point()
bounding_boxes = bounding_boxes.float() if need_cast else bounding_boxes.clone()
bounding_boxes = (
convert_bounding_box_format(bounding_boxes, old_format=format, new_format=intermediate_format)
convert_bounding_box_format(bounding_boxes, old_format=format, new_format=intermediate_format, inplace=True)
).reshape(-1, 8 if is_rotated else 4)

dtype = bounding_boxes.dtype if torch.is_floating_point(bounding_boxes) else torch.float32
dtype = bounding_boxes.dtype
device = bounding_boxes.device

# perspective_coeffs are computed as endpoint -> start point
Expand Down Expand Up @@ -2430,18 +2431,21 @@ def elastic_bounding_boxes(

# TODO: add in docstring about approximation we are doing for grid inversion
device = bounding_boxes.device
dtype = bounding_boxes.dtype if torch.is_floating_point(bounding_boxes) else torch.float32
original_dtype = bounding_boxes.dtype
is_rotated = tv_tensors.is_rotated_bounding_format(format)

original_shape = bounding_boxes.shape
need_cast = not bounding_boxes.is_floating_point()
bounding_boxes = bounding_boxes.float() if need_cast else bounding_boxes.clone()
dtype = bounding_boxes.dtype

if displacement.dtype != dtype or displacement.device != device:
displacement = displacement.to(dtype=dtype, device=device)

original_shape = bounding_boxes.shape
# TODO: first cast to float if bbox is int64 before convert_bounding_box_format
intermediate_format = tv_tensors.BoundingBoxFormat.CXCYWHR if is_rotated else tv_tensors.BoundingBoxFormat.XYXY

bounding_boxes = (
convert_bounding_box_format(bounding_boxes.clone(), old_format=format, new_format=intermediate_format)
convert_bounding_box_format(bounding_boxes, old_format=format, new_format=intermediate_format, inplace=True)
).reshape(-1, 5 if is_rotated else 4)

id_grid = _create_identity_grid(canvas_size, device=device, dtype=dtype)
Expand Down Expand Up @@ -2473,10 +2477,12 @@ def elastic_bounding_boxes(
out_bboxes, format=intermediate_format, canvas_size=canvas_size, clamping_mode=clamping_mode
)

return convert_bounding_box_format(
out_bboxes, old_format=intermediate_format, new_format=format, inplace=False
out_bboxes = convert_bounding_box_format(
out_bboxes, old_format=intermediate_format, new_format=format, inplace=True
).reshape(original_shape)

return out_bboxes.to(original_dtype)


@_register_kernel_internal(elastic, tv_tensors.BoundingBoxes, tv_tensor_wrapper=False)
def _elastic_bounding_boxes_dispatch(
Expand Down