[Feature] Support CopyPaste for RotatedBoxes

[nijkah]

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

Closes #82

Modification

1. Add RBitMasks
2. Add RCopyPaste

Welcome any suggestions on the current implementation!

Example

[CLAassistant]

All committers have signed the CLA.

[nijkah]

Here, we need to support mmrotate.core.bbox.poly2obb_np for various angle versions.
However, poly2obb supposes that polygons have 4 vertices.
How can we support this?

[zytx121]

In mmrotate1.x, we suppose all rotated boxes use oc angle version during data transforms.
We only need to distinguish different angle versions within the box_head.
So I think it's OK to write like this.

[zytx121]

Excellent job! It's very useful.

[zytx121]

@nijkah Sorry to reply so late. Anyway, I'm back.

The project structure of MMDet and mmrotate has been updated. Please update your code according to https://github.com/open-mmlab/mmdetection/blob/3.x/mmdet/structures/mask/structures.py

I will continue to pay attention to this important PR and hope your enthusiasm will not fade.

[hongxrsysu]

I use this code, and met a bug:

File "tools/misc/browse_dataset.py", line 110, in
main()
File "tools/misc/browse_dataset.py", line 86, in main
for item in dataset:
File "/root/miniconda3/lib/python3.8/site-packages/mmdet/datasets/dataset_wrappers.py", line 431, in getitem
updated_results = transform(copy.deepcopy(results))
File "/root/miniconda3/lib/python3.8/site-packages/mmdet/datasets/pipelines/transforms.py", line 2830, in call
selected_results = self._select_object(results['mix_results'][0])
File "/root/miniconda3/lib/python3.8/site-packages/mmdet/datasets/pipelines/transforms.py", line 2839, in _select_object
masks = results['gt_masks']
KeyError: 'gt_masks'

I find that "gt_masks" is in the "mmrotate/datasets/pipelines/transforms.py".

So how to fix it?

[nijkah]

@hongxrsysu I think it works normally after upgrading mmdet >= 2.26.0.
It supposes that CopyPaste supports generating a pseudo gt_masks when gt_masks is not available.

Please refer to these links.
https://github.com/open-mmlab/mmdetection/releases/tag/v2.26.0
open-mmlab/mmdetection#8905

[nijkah]

@zytx121 The CopyPaste implementation seems quite different between mmdet==2.x and mmdet==3.x.
For example, in mmdet==3.x, it does not support generating pseudo masks from bounding boxes.

Is there any good way to support both mmrotate==1.x and mmrotate==0.x?
Or would it be okay to implement each in a different branch?

[hongxrsysu]

@hongxrsysu I think it works normally after upgrading mmdet >= 2.26.0. It supposes that CopyPaste supports generating a pseudo gt_masks when gt_masks is not available.

Please refer to these links. https://github.com/open-mmlab/mmdetection/releases/tag/v2.26.0 open-mmlab/mmdetection#8905

Thanks! It works!

[zytx121]

@zytx121 The CopyPaste implementation seems quite different between mmdet==2.x and mmdet==3.x. For example, in mmdet==3.x, it does not support generating pseudo masks from bounding boxes.

Is there any good way to support both mmrotate==1.x and mmrotate==0.x? Or would it be okay to implement each in a different branch?

Hi @nijkah!
As far as I know, there is currently no way to support two branches at the same time. It is okay to implement each in a different branch. Due to a shortage of personnel, mmrotate0. x has actually stopped maintenance since 2023. So the priority of 1. x will increase.

In 1. x, projects\ provides a quick iterative code solution that does not require strict unit testing and documentation.

[nijkah]

Hi, @zytx121. Apologize for the slow progress.
Recently. I create a PR #877 for MMRotate=1.x. 😄

I'll close this PR.