Rotated bboxes transforms by AntoineSimoulin · Pull Request #9084 · pytorch/vision

AntoineSimoulin · 2025-05-26T03:11:38Z

Add Transforms support for Rotated Boxes

This PR is the first of a series to add support for rotated boxes transforms. In particular this PR implements the following functionalities:

Add support to draw rotated boxes by modifying the function draw_bounding_boxes to accept BoundingBoxes with format BoundingBoxFormat.XYXYXYXY;
Add utility function is_rotated_bounding_format to infer whether a BoundingBoxFormat corresponds to a rotated format or not. To preserve torchscript support, this function is not added as a method from the Enum definition of BoundingBoxFormat;
Create function to clamp rotated boxes. For now the function only return a clone of the input boxes. The clamping function is applied after each transform. To pass the test we created the function while the functionalities are not implemented yet;
Add support for vertical and horizontal flip for rotated boxes.

Test plan

Please run the following tests:

pytest test/test_tv_tensors.py -vvv -k "test_bbox_format"
pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_kernel_bounding_boxes"
pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_bounding_boxes_correctness"
pytest test/test_transforms_v2.py -vvv -k "TestVerticalFlip and test_kernel_bounding_boxes"
pytest test/test_transforms_v2.py -vvv -k "TestVerticalFlip and test_bounding_boxes_correctness"

Plot function

The plot function can be used as follow

import torch
from gallery.transforms.helpers import plot
from test.common_utils import make_bounding_boxes
from torchvision import transforms, tv_tensors

img = torch.ones(3, 360, 360)
boxes = make_bounding_boxes(
    canvas_size=(360, 360),
    format=tv_tensors.BoundingBoxFormat.XYXYXYXY,
    num_boxes=2,
)
plot([(img, boxes)])

Future work

This PR implements only two transforms for rotated boxes. However, it is intended to validate the implementation before releasing other transforms. Please also note that the clamping transforms is for now just a placeholder to make sure we do pass the tests for rotated boxes.

cc @vfdev-5

Test Plan: Run unit tests:`pytest test/test_tv_tensors.py -vvv -k "test_bbox_format"`

Test Plan: Run unit tests: `pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_kernel_bounding_boxes"` and `pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_bounding_boxes_correctness"`

Test Plan: Run unit tests: `pytest test/test_transforms_v2.py -vvv -k "TestVerticalFlip and test_kernel_bounding_boxes"`

pytorch-bot · 2025-05-26T03:11:41Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9084

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

Experiencing "429: Too Many Requests" on downloading actions

This comment was automatically generated by Dr. CI and updates every 15 minutes.

NicolasHug

Thanks for the PR @AntoineSimoulin , I made a first round of comments

torchvision/utils.py

NicolasHug · 2025-05-29T10:22:47Z

torchvision/utils.py

    elif image.size(0) not in {1, 3}:
        raise ValueError("Only grayscale and RGB images are supported")
-    elif (boxes[:, 0] > boxes[:, 2]).any() or (boxes[:, 1] > boxes[:, 3]).any():
+    elif boxes.shape[-1] == 4 and ((boxes[:, 0] > boxes[:, 2]).any() or (boxes[:, 1] > boxes[:, 3]).any()):


Nit: since we seem to use both boxes.shape[-1] == 4 and len(bbox) == 4 as condition checks, it might help to create a unified boolean variable e.g.

is_rotated = boxes.shape[-1] == 4

and use is_rotated for the remainder of the function. It also makes the conditions slightly more explicit about what they're checking against.

I would prefer not to create a variable before the testing at the beginning of the function. But I simplified the core of the function given your other comments!

torchvision/utils.py

torchvision/transforms/v2/functional/_geometry.py

test/test_transforms_v2.py

NicolasHug · 2025-05-29T11:31:49Z

test/test_transforms_v2.py

+            # operation
+            dtype = output.dtype
+
+        return output.to(dtype=dtype, device=device)


Shouldn't we cast back to the input dtype unconditionally? In general the transforms should preserve the input dtype, but here's it's not clear that we are?

In this test, we are creating an intermediate tensor and therefore make sure to cast it to the correct dtype.

AntoineSimoulin

@NicolasHug thanks for reviewing! I should have addressed all comments here. Additional tests for visualization can be run with pytest test/test_utils.py -vvv -k "test_draw_rotatated_boxes". Let me know if you have additional comment to add in this review.

AntoineSimoulin · 2025-05-29T19:51:11Z

test/test_transforms_v2.py

+            # operation
+            dtype = output.dtype
+
+        return output.to(dtype=dtype, device=device)


In this test, we are creating an intermediate tensor and therefore make sure to cast it to the correct dtype.

AntoineSimoulin · 2025-05-29T19:52:45Z

torchvision/tv_tensors/_bounding_boxes.py


+# TODO: Once torchscript supports Enums with staticmethod
+# this can be put into BoundingBoxFormat as staticmethod
+def is_rotated_bounding_format(format: BoundingBoxFormat) -> bool:


Hey @NicolasHug, I tried to do that. Unfortunately it is not compatible with TorchScript. Indeed BoundingBoxes is directly inheriting from torch.Tensor and TorchScript does not fully support inheritance from built-in PyTorch types like torch.Tensor and has specific rules and limitations regarding which methods and attributes are accessible when scripting custom classes that inherit from these types.

torchvision/utils.py

AntoineSimoulin · 2025-05-29T19:53:54Z

torchvision/utils.py

-        boxes (Tensor): Tensor of size (N, 4) containing bounding boxes in (xmin, ymin, xmax, ymax) format. Note that
-            the boxes are absolute coordinates with respect to the image. In other words: `0 <= xmin < xmax < W` and
-            `0 <= ymin < ymax < H`.
+        boxes (Tensor): Tensor of size (N, 4) or (N, 8) containing bounding boxes.


This sounds good. I propose we do address this in a subsequent PR!

AntoineSimoulin · 2025-05-29T19:54:31Z

torchvision/utils.py

    elif image.size(0) not in {1, 3}:
        raise ValueError("Only grayscale and RGB images are supported")
-    elif (boxes[:, 0] > boxes[:, 2]).any() or (boxes[:, 1] > boxes[:, 3]).any():
+    elif boxes.shape[-1] == 4 and ((boxes[:, 0] > boxes[:, 2]).any() or (boxes[:, 1] > boxes[:, 3]).any()):


I would prefer not to create a variable before the testing at the beginning of the function. But I simplified the core of the function given your other comments!

NicolasHug

Thank you @AntoineSimoulin !

test/test_transforms_v2.py

github-actions · 2025-06-04T02:38:47Z

Hey @AntoineSimoulin!

You merged this PR, but no labels were added.
The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py

Summary: Co-authored-by: Nicolas Hug <contact@nicolas-hug.com> Reviewed By: AntoineSimoulin Differential Revision: D79175063 fbshipit-source-id: 7389510a020fae499cfe8812a0d26d9adb90c744

AntoineSimoulin added 5 commits May 25, 2025 17:02

Add support for rotated boxes in draw_bounding_boxes

734aed2

Add utility function to identify rotated box formats

9827ab6

Test Plan: Run unit tests:`pytest test/test_tv_tensors.py -vvv -k "test_bbox_format"`

Modify clamping for rotated boxes

87a238c

Add horizontal_flip_rotated_bounding_boxes

95ed7cf

Test Plan: Run unit tests: `pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_kernel_bounding_boxes"` and `pytest test/test_transforms_v2.py -vvv -k "TestHorizontalFlip and test_bounding_boxes_correctness"`

Add vertical_flip_bounding_boxes

a7d07dc

Test Plan: Run unit tests: `pytest test/test_transforms_v2.py -vvv -k "TestVerticalFlip and test_kernel_bounding_boxes"`

facebook-github-bot added the cla signed label May 26, 2025

AntoineSimoulin added 3 commits May 27, 2025 15:17

Add visualization for rotated boxes

3996daa

Fix vertical flip orientation

3b4100c

Fix horizontal flip orientation

e223c6f

NicolasHug reviewed May 29, 2025

View reviewed changes

Fix PR comments for rorated box transforms

36b02dd

AntoineSimoulin commented May 29, 2025

View reviewed changes

NicolasHug added 2 commits May 30, 2025 16:54

use tensor instead of Tensor which is the type

57f2452

Add non-parallel bbox

a15a057

NicolasHug approved these changes May 30, 2025

View reviewed changes

test/test_transforms_v2.py Outdated Show resolved Hide resolved

Merge branch 'main' into rotated-bboxes-transforms

4bde5e5

AntoineSimoulin merged commit 428a54c into pytorch:main Jun 4, 2025
39 of 41 checks passed

AntoineSimoulin deleted the rotated-bboxes-transforms branch June 4, 2025 03:00

AntoineSimoulin mentioned this pull request Jun 4, 2025

More rotated bboxes transforms #9095

Merged

NicolasHug added module: transforms new feature labels Jun 4, 2025

AntoineSimoulin mentioned this pull request Jun 10, 2025

Rotated bboxes transforms #9104

Merged

Conversation

AntoineSimoulin commented May 26, 2025 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Add Transforms support for Rotated Boxes

Test plan

Plot function

Future work

Uh oh!

pytorch-bot bot commented May 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9084

❗ 1 Active SEVs

Uh oh!

NicolasHug left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AntoineSimoulin left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

NicolasHug left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Jun 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

AntoineSimoulin commented May 26, 2025 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented May 26, 2025 •

edited

Loading