[Feature] Distribution Angle Coder (WIP)

[liuyanyi]

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

Introduce the distribution representation for angle to achieve a differentiable angular encoding, which is used in GFL model for bbox representation.

However, there are still problems of angular boundaries and double peaks of square-like targets.

Model can converge on HRSC, 6x schedule with RR.

10/13 14:43:21 - mmengine - INFO - Epoch(val) [72][453/453]  dota/mAP: 0.6480  dota/AP50: 0.8950  dota/AP55: 0.8940  dota/AP60: 0.8930  dota/AP65: 0.8900  dota/AP70: 0.8770  dota/AP75: 0.7800  dota/AP80: 0.6500  dota/AP85: 0.4220  dota/AP90: 0.1550  dota/AP95: 0.0230

Support:

• [x] Rotated FCOS
• [ ] AngleBranchRetinaHead
• [ ] Future Models...

For Distribution Angle Coder, the encode size is 17, but target encode size is 1, so AngleBranchRetinaHead should modify the logic of get angle target.

Modification

Please briefly describe what modification is made in this PR.

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. The documentation has been modified accordingly, like docstring or example tutorials.