fix bbox direction

[zhulf0804]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

When training PointPillars, the bbox direction gt label is set to 0 when the angle is between in [0, np.pi] and is set to 1 when the angle is between in [np.pi, 2*np.pi].

mmdetection3d/mmdet3d/models/dense_heads/train_mixins.py

Lines 336 to 338 in ed81c6c

	offset_rot = limit_period(rot_gt - dir_offset, 0, 2 * np.pi)
	dir_cls_targets = torch.floor(offset_rot / (2 * np.pi / num_bins)).long()
	dir_cls_targets = torch.clamp(dir_cls_targets, min=0, max=num_bins - 1)

However, in the test stage, the operation may be opposite. That is the predicted label 0 corresponds to [-np.pi, 0] or [np.pi, 2*np.pi], while the predicted label 1 corresponds to [0, np.pi]

mmdetection3d/mmdet3d/models/dense_heads/anchor3d_head.py

Lines 506 to 511 in ed81c6c

	if bboxes.shape[0] > 0:
	dir_rot = limit_period(bboxes[..., 6] - self.dir_offset,
	self.dir_limit_offset, np.pi)
	bboxes[..., 6] = (
	dir_rot + self.dir_offset +
	np.pi * dir_scores.to(bboxes.dtype))

Modification

I modified in predicted bbox prediction (mmdet3d/models/dense_heads/anchor3d_head.py#L506-L511) as the following code:

if bboxes.shape[0] > 0:
            dir_rot = limit_period(bboxes[..., 6] - self.dir_offset,
                                   self.dir_limit_offset, np.pi)
            bboxes[..., 6] = (
                dir_rot + self.dir_offset + np.pi *
                (1 - dir_scores.to(bboxes.dtype)))

BC-breaking (Optional)

Does the modification introduce changes that break the back-compatibility of the downstream repos?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[CLAassistant]

All committers have signed the CLA.

[ZwwWayne]

Hi @zhulf0804 ,
Thanks for your kind PR. Could you fix the lint issue? You can find more details here.

[zhulf0804]

Thank you for your approval. I'll fix it as soon as possible. Best wishes. Lifa. Update: Fixed it. Is it ok now ?

[Tai-Wang]

This bug seems to be fixed in the latest mmdet3d. I found your used version is about v0.16.0. This bug can be fixed via setting the dir_limit_offset to 0 to keep it consistent with that in get_direction_target. You can also refer to this issue in SECOND for more details.

[Tai-Wang]

Please double-check whether this problem exists in the latest mmdet3d. I guess it has been fixed as I commented.

[zhulf0804]

Thank you for your quick reply. I checked that It has been fixed in both master and v1.0.0rc0.
However, I just used the previous version v0.18.1.

I thought I used a very new version, because v0.18.1 was released on Feb 09, 2022. I'm sorry for not checking the latest version before asking the question.

By the way, I found the bug was fixed in #803 on Jul 29, 2021.
Then It was not updated until recently in v1.0.0rc0 version 22 days ago (maybe on Mar 1, 2022.). I wonder why not considering updating it in earlier versions ?

Looking forward to your reply.

Anyway, nice repo and implementations.