[Feat]: Support DAB-DETR

[KeiChiTse]

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

Implementation of DAB-DETR detector.

Modification

Add DAB-DETR detector module and DAB-DETR head module, main modifications are in mmdet/models/utils/transformer.py, mmdet/models/dense_heads/dab_detr_head.py files.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-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)

Use DAB-DETR detector.

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

[CLAassistant]

All committers have signed the CLA.

[KeiChiTse]

20220814_041356.log
20220816_015156.log

[Update]:

Experiment logs of "res50" and "res50+3 patterns" are uploaded. Compared to paper, as for res50, current mAP is 41.8 (-0.4) with 80 classes. As for res50+3 patterns, current mAP is 42.7 (+0.1) with 80 classes.

The experiment results are supported by only one individual run for each setting (for now), to improve the mAP, more experiments and further investigation are WIP!

[KeiChiTse]

[Update]:

Previous experiments set dropout to 0.1, misaligned with official repo as they use dropout 0.0 , which is proved to be beneficial in enhancing mAP in Sparse R-CNN paper.
More experiments with dropout 0.0 and num classes 80 are on the way!

[KeiChiTse]

[Update]:

Pretrained checkpoint of this version will be available after the refactor work is done in upcoming MMDetection 3.0.
For details, please refer to #8811 and #8763.