[Feature] add Mask2former to mmdet

[chhluo]

This PR Mask2former is based on maskformer's PR(#6607 ).The components of maskformer in mask2former are same with the origin PR(#6607 ).

Motivation

Add a greater panoptic segmentation model to mmdet.

Modification

• Components in maskformer(not list here).
• Multi-scale-Deformable-Attention Pixeldecoder.
• mask2former head and detector.
• two configs for r50 and swin-tiny.
• instance segmentation evaluation in CocoPanopticDataset.

evaluation result of mask2former_r50_lsj_8x2_50e_coco
{
'PQ': 51.85113493826705, 'SQ': 83.07220152800042, 'RQ': 61.59141241640433, 'PQ_th': 57.712003700037954, 'SQ_th': 84.06726498653896, 'RQ_th': 68.07799535626707, 'PQ_st': 43.00454058087707, 'SQ_st': 81.57021894907417, 'RQ_st': 51.80034382793225, 
'PQ_copypaste': '51.9 83.1 61.6 57.7 84.1 68.1 43.0 81.6 51.8', 
'bbox_mAP': 0.443, 'bbox_mAP_50': 0.637, 'bbox_mAP_75': 0.477, 'bbox_mAP_s': 0.264, 'bbox_mAP_m': 0.475, 'bbox_mAP_l': 0.632, 
'bbox_mAP_copypaste': '0.443 0.637 0.477 0.264 0.475 0.632', 
'segm_mAP': 0.417, 'segm_mAP_50': 0.64, 'segm_mAP_75': 0.449, 'segm_mAP_s': 0.223, 'segm_mAP_m': 0.454, 'segm_mAP_l': 0.638, 
'segm_mAP_copypaste': '0.417 0.640 0.449 0.223 0.454 0.638'
}

• processing for instance segmengtation results in test.py.
• More general version of match cost: DiceCost, MaskFocalLossCost, and the new match cost: CrossEntropyLossCost.
• bce loss for input and pred which have same shape.

[codecov]

Codecov Report

Merging #6938 (c9ba372) into dev (1a9087b) will increase coverage by 0.50%.
The diff coverage is 93.20%.

❗ Current head c9ba372 differs from pull request most recent head 2ad5723. Consider uploading reports for the commit 2ad5723 to get more accurate results

@@            Coverage Diff             @@
##              dev    #6938      +/-   ##
==========================================
+ Coverage   63.62%   64.13%   +0.50%     
==========================================
  Files         341      345       +4     
  Lines       27585    27928     +343     
  Branches     4660     4699      +39     
==========================================
+ Hits        17552    17911     +359     
+ Misses       9109     9079      -30     
- Partials      924      938      +14     

Flag	Coverage Δ
unittests	64.10% <93.20%> (+0.50%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet/datasets/coco.py	58.22% <ø> (ø)
...ds/panoptic_fusion_heads/maskformer_fusion_head.py	75.82% <75.82%> (ø)
mmdet/models/detectors/maskformer.py	68.57% <95.83%> (+5.41%)	⬆️
mmdet/models/dense_heads/mask2former_head.py	96.68% <96.68%> (ø)
mmdet/models/plugins/msdeformattn_pixel_decoder.py	99.00% <99.00%> (ø)
mmdet/core/bbox/match_costs/__init__.py	100.00% <100.00%> (ø)
mmdet/core/bbox/match_costs/match_cost.py	97.97% <100.00%> (+0.51%)	⬆️
mmdet/core/mask/__init__.py	100.00% <100.00%> (ø)
mmdet/core/mask/utils.py	64.86% <100.00%> (+16.86%)	⬆️
mmdet/datasets/coco_panoptic.py	90.07% <100.00%> (+0.07%)	⬆️
... and 15 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fc8fb16...2ad5723. Read the comment docs.

[PeterVennerstrom]

Are there plans to support instance segmentation only with mask2former? I trained an instance segmentation mask2former with swin t 50e on COCO and got close to the published performance.

panoptic_gt_processing.py pads the instance seg gt masks using the shape from the semantic seg gt masks. I made gt_semantic_seg an optional argument in forward_train() and wrote a preprocess_instance_gt() method that used pad_shape from img_metas rather than taking the shape from gt_semantic_seg. These were the only non-config changes.

My approach isn't a good design.

[chhluo]

Are there plans to support instance segmentation only with mask2former? I trained an instance segmentation mask2former with swin t 50e on COCO and got close to the published performance.

panoptic_gt_processing.py pads the instance seg gt masks using the shape from the semantic seg gt masks. I made gt_semantic_seg an optional argument in forward_train() and wrote a preprocess_instance_gt() method that used pad_shape from img_metas rather than taking the shape from gt_semantic_seg. These were the only non-config changes.

My approach isn't a good design.

Good suggestion, we will add this proposal to our backlog.
If you are interested in implementing it, you can propose a PR, we can do it together !

[PeterVennerstrom]

Good suggestion, we will add this proposal to our backlog. If you are interested in implementing it, you can propose a PR, we can do it together !

Excellent, thanks for your work on this, it's a really nice implementation.

I've got a pull request out which changes the gt_preprocessing logic to use pad_shape from img_metas and handle optional gt_semantic_seg. This enables forward_train() using the standard coco.py dataset used for instance segmentation.

Assuming this is acceptable, it leaves the matter of evaluation. We could adjust the detector to format the results differently depending the the dataset type (panoptic or instance seg) used to train or we could adjust coco.py to parse the new PQ format results (which would only contain instance seg and bbox results).