mmdet: r3_cnn original code

[hachreak]

Signed-off-by: Leonardo Rossi leonardo.rossi@unipr.it

Motivation

Original code for the paper: Recursively Refined R-CNN: Instance Segmentation with Self-RoI Rebalancing

Modification

R3RoIHead is a more generic version of HybridTaskCascadeRoIHead, in the sense that permits to specify which bbox/mask head use at each step.

BC-breaking (Optional)

It should not break anything.

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

Merging #6399 (755e03c) into master (6cf9aa1) will decrease coverage by 0.06%.
The diff coverage is 27.29%.

❗ Current head 755e03c differs from pull request most recent head 4c2e563. Consider uploading reports for the commit 4c2e563 to get more accurate results

@@            Coverage Diff             @@
##           master    #6399      +/-   ##
==========================================
- Coverage   62.03%   61.97%   -0.07%     
==========================================
  Files         319      321       +2     
  Lines       25314    25386      +72     
  Branches     4189     4211      +22     
==========================================
+ Hits        15703    15732      +29     
- Misses       8783     8824      +41     
- Partials      828      830       +2     

Flag	Coverage Δ
unittests	61.94% <27.29%> (-0.07%)	⬇️

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

Impacted Files	Coverage Δ
mmdet/models/roi_heads/r3_roi_head.py	25.35% <25.35%> (ø)
mmdet/models/roi_heads/mask_heads/r3_mask_head.py	27.58% <27.58%> (ø)
mmdet/models/roi_heads/cascade_roi_head.py	36.72% <66.66%> (-0.21%)	⬇️
mmdet/models/roi_heads/__init__.py	100.00% <100.00%> (ø)
mmdet/models/roi_heads/htc_roi_head.py	100.00% <100.00%> (+71.38%)	⬆️
mmdet/models/roi_heads/mask_heads/__init__.py	100.00% <100.00%> (ø)
mmdet/models/backbones/pvt.py	83.59% <0.00%> (-0.45%)	⬇️
mmdet/apis/__init__.py	100.00% <0.00%> (ø)
mmdet/models/backbones/swin.py	81.63% <0.00%> (+0.06%)	⬆️
mmdet/apis/train.py	16.00% <0.00%> (+0.09%)	⬆️
... and 4 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 6cf9aa1...4c2e563. Read the comment docs.

[ZwwWayne]

Hi @hachreak ,
Thanks for your kind contribution. Would you like to fix the lint issue first? You can click the buttons in this section to see the details.

[hachreak]

Would you like to fix the lint issue first?

Hi @ZwwWayne , thanks for your comment!
I rebased and I hope to have fixed the issues now (I run pre-commit run --all-files). 😄

[hhaAndroid]

@hachreak Please add the unit test, because the code coverage drops more.

[hhaAndroid]

Like other algorithms, the precision table needs to be added.

[hhaAndroid]

@hachreak Please confirm whether the Cascade Mask RCNN has BC breaking. At the same time, you can simply run an experiment to see if there are any problems. If you need other help, you can contact me.

[hachreak]

Hi @hhaAndroid
I think it should not break anything for Cascade R-CNN and neither for HTC.
But I did't consider Cascade Mask RCNN.
Where is the source code for that model to check?

Actually, to run experiments for me now could be a problem because I don't have enough gpus to start some training. 😞

[hachreak]

Hi @hhaAndroid any news for the PR? 😄
How could I help to speed up the integration?

[ZwwWayne]

1. Conflicts should be resolved.
2. htc_roi_head should not be deleted.
3. cascade_roi_head.py should not be modified. We suggest restricting the modification inside the r3_roi_head.py You can override these functions in cascade_roi_head.py

[hachreak]

Hi @ZwwWayne
thanks to answering 😄

1. of course! I'll do it
2. actually I'm not deleting but generalize it, because R3 RoI Head is a generalization of HTC RoI Head, where I can select the BBox Head to use each time. I could implement a R3 RoI Head separately but it means doubling 90% of the code. Do you think I should reimplement anyway everything? Maybe I can find another way to do it. let me think about it 😄
3. I needed to make small changes (but without affecting the Cascade behaviour) to permit me to implement the loop. I mean, I needed to be able to select the BBox Head I want during each step of the for loop (the for loop which implement the cascade, and in my case the loop).
   Basically, for Cascade it is mandatory to select the next one, in my case I configure it.