[Fix] Discard bboxes with sizes equals to min_bbox_size

[yeliudev]

As described in #5005, sometimes anchor heads may produce proposals with area == 0, which brings errors during mask prediction (masks can not be predicted when the areas of bboxes are 0). A possible solution is setting min_bbox_size to a value larger than 0, but I suggest simply discarding all the proposals with area == min_bbox_size in anchor heads directly.

[codecov]

Codecov Report

Merging #5011 (c6e1f5e) into master (911eb11) will increase coverage by 0.19%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #5011      +/-   ##
==========================================
+ Coverage   65.48%   65.68%   +0.19%     
==========================================
  Files         257      257              
  Lines       19987    20080      +93     
  Branches     3419     3419              
==========================================
+ Hits        13088    13189     +101     
+ Misses       6186     6178       -8     
  Partials      713      713              

Flag	Coverage Δ
unittests	65.64% <0.00%> (+0.19%)	⬆️

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

Impacted Files	Coverage Δ
mmdet/models/dense_heads/cascade_rpn_head.py	24.34% <0.00%> (ø)
mmdet/models/dense_heads/ga_rpn_head.py	27.27% <0.00%> (ø)
mmdet/models/dense_heads/rpn_head.py	76.47% <0.00%> (+5.88%)	⬆️
mmdet/models/losses/ae_loss.py	100.00% <0.00%> (ø)
mmdet/models/losses/mse_loss.py	100.00% <0.00%> (ø)
mmdet/models/losses/smooth_l1_loss.py	100.00% <0.00%> (ø)
mmdet/core/bbox/coder/base_bbox_coder.py	100.00% <0.00%> (ø)
mmdet/models/roi_heads/mask_heads/maskiou_head.py	100.00% <0.00%> (ø)
.../models/roi_heads/mask_heads/feature_relay_head.py	100.00% <0.00%> (ø)
...models/roi_heads/mask_heads/fused_semantic_head.py	100.00% <0.00%> (ø)
... and 36 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 911eb11...c6e1f5e. Read the comment docs.

[ZwwWayne]

Hi @yeliudev ,
Thanks for the PR. Could you verify the testing performance of related models on the COCO dataset?
Also, since it modifies RPN head and may affect ONNX exportation, could you also check whether it could export Faster RCNN ONNX models correctly in this PR? You can follow this doc.

[yeliudev]

I've tested the following models and all the results remain unchanged. The ONNX export also works correctly because this PR does not change its behaviors.

Model	Original mAP	New mAP
Faster R-CNN	37.4	37.4
Libra R-CNN	38.3	38.3
Cascade R-CNN	40.3	40.3
Mask R-CNN	38.2 / 34.7	38.2 / 34.7
Cascade Mask R-CNN	41.2 / 35.9	41.2 / 35.9
PointRend	38.4 / 36.3	38.4 / 36.3

[ZwwWayne]

This PR is related to post-processing. FYI @jshilong