[Feature] Support DGCNN

[DCNSW]

Motivation

Support DGCNN (Dynamic Graph CNN for Learning on Point Clouds) in mmdetection3d.

Modification

Design new configs, backbones, decode_heads and ops for DGCNN.

BC-breaking (Optional)

Does the modification introduce changes that break the back-compatibility of the downstream repos?
No.

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

[codecov]

Codecov Report

Merging #801 (18541d9) into master (cab70db) will increase coverage by 0.14%.
The diff coverage is 63.52%.

@@            Coverage Diff             @@
##           master     #801      +/-   ##
==========================================
+ Coverage   49.32%   49.47%   +0.14%     
==========================================
  Files         210      216       +6     
  Lines       16019    16179     +160     
  Branches     2556     2574      +18     
==========================================
+ Hits         7902     8004     +102     
- Misses       7614     7670      +56     
- Partials      503      505       +2     

Flag	Coverage Δ
unittests	49.47% <63.52%> (+0.14%)	⬆️

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

Impacted Files	Coverage Δ
mmdet3d/ops/dgcnn_modules/dgcnn_fa_module.py	52.17% <52.17%> (ø)
mmdet3d/ops/dgcnn_modules/dgcnn_gf_module.py	55.93% <55.93%> (ø)
mmdet3d/models/decode_heads/dgcnn_head.py	60.00% <60.00%> (ø)
mmdet3d/ops/dgcnn_modules/dgcnn_fp_module.py	64.70% <64.70%> (ø)
mmdet3d/models/backbones/dgcnn.py	78.12% <78.12%> (ø)
mmdet3d/models/backbones/__init__.py	100.00% <100.00%> (ø)
mmdet3d/models/decode_heads/__init__.py	100.00% <100.00%> (ø)
mmdet3d/ops/__init__.py	100.00% <100.00%> (ø)
mmdet3d/ops/dgcnn_modules/__init__.py	100.00% <100.00%> (ø)
mmdet3d/models/fusion_layers/point_fusion.py	64.44% <0.00%> (-0.78%)	⬇️
... and 10 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 cab70db...18541d9. Read the comment docs.

[Wuziyi616]

The code needs to be refined. Benchmark is required. I think the design of DGCNN is acceptable (via ops/dgcnn_modules), but we need to polish the code. Also, unit tests are needed. See my comments above.

[Wuziyi616]

Let's wait for results on 6-fold evaluation.

[Wuziyi616]

@ZwwWayne Please have a look here. The benchmark is done on S3DIS (both Area 5 and 6-Fold). The code in each added module is good. However, I don't know if you think the overall design of DGCNN is acceptable (I think it's okay). Similar to PointNet++, in this repo, DGCNN is decomposed to several sub-modules, including GraphFeature (gf), FeatureAggregation (fa) and FeaturePropagation (fp). Please kindly have a look.

[Wuziyi616]

@DCNSW Please kindly check the docstring of the code in this PR. Add optional to those args which has default values.

[Wuziyi616]

I think this PR is almost ready

[Wuziyi616]

After all the comments are resolved, please use this script to gather trained model weight and zip it and send it to @ZwwWayne.

[Wuziyi616]

Great job. Cheers!

[ZwwWayne]

Considering reduce the registry of fa/fg module

[Wuziyi616]

Some benchmark release related modifications are missed. Please refer to #541 for more details, e.g. adding DGCNN model in docs/model_zoo.md, README.md and README_zh-CN.md files

[Tai-Wang]

Please also fix minor problems similar to our comments