[Feature] Support S3DIS data pre-processing and dataset class

[Wuziyi616]

• Provide instructions to S3DIS data downloading
• Provide scripts and meta_data for S3DIS data pre-processing
• Add a S3DISSegDataset class
• Add a config file for S3DIS data loading

[Wuziyi616]

Different from other datasets like ScanNet, S3DIS doesn't have a train-val split! S3DIS has 6 areas of data (can think of them as 6 subsets, collected in different buildings/regions). The common setting is training on 5 of them, and testing on the remaining one. People often report an Area_5 test result, and an average result. The average result means the average mIoU of 6 experiments (leaving Area_1/2/.../6 respectively).

Therefore, in create_data.py, I generate one info file for one area, so that we can concat 5 to form a training dataset and use the remaining one for test dataset.

In s3dis_dataset.py, I write an inner dataset class (_S3DISSegDataset) for data loading on 1 area, and further write a wrapper class (S3DISSegDataset) to hold (concat) multiple areas.

[codecov]

Codecov Report

Merging #433 (3eff65b) into master (78c29c3) will increase coverage by 0.26%.
The diff coverage is 90.80%.

❗ Current head 3eff65b differs from pull request most recent head bfadf26. Consider uploading reports for the commit bfadf26 to get more accurate results

@@            Coverage Diff             @@
##           master     #433      +/-   ##
==========================================
+ Coverage   51.04%   51.31%   +0.26%     
==========================================
  Files         187      188       +1     
  Lines       13797    13884      +87     
  Branches     2238     2259      +21     
==========================================
+ Hits         7043     7124      +81     
- Misses       6277     6280       +3     
- Partials      477      480       +3     

Flag	Coverage Δ
unittests	51.31% <90.80%> (+0.26%)	⬆️

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

Impacted Files	Coverage Δ
mmdet3d/datasets/custom_3d_seg.py	65.66% <ø> (ø)
mmdet3d/datasets/scannet_dataset.py	94.56% <ø> (ø)
mmdet3d/datasets/s3dis_dataset.py	90.69% <90.69%> (ø)
mmdet3d/datasets/__init__.py	100.00% <100.00%> (ø)
mmdet3d/datasets/pipelines/transforms_3d.py	86.92% <0.00%> (+0.43%)	⬆️

---

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 2c29afa...bfadf26. Read the comment docs.

[Tai-Wang]

Please check all the docstring, there are still many nonstandard formats.
Next time we can split similar PR into several small ones.
It is so heavy to review and merge with so many things in the single one 0.0

[Wuziyi616]

I just copy-paste the pre-processing code of S3DIS from other repo so it's dirty, I am sorry. I check it and find many unnecessary functions. I will remove and make the code cleaner.

Yes, next time I will split into small PRs. I apologize for that.

[Tai-Wang]

I just copy-paste the pre-processing code of S3DIS from other repo so it's dirty, I am sorry. I check it and find many unnecessary functions. I will remove and make the code cleaner.

Yes, next time I will split into small PRs. I apologize for that.

OK, it doesn't matter. You have done a good job :)

[Wuziyi616]

Need to add pipeline support in S3DISSegDataset.show/evaluate() after PR#430 is merged.

[Wuziyi616]

Ready for review now @Tai-Wang @ZwwWayne.

[Wuziyi616]

Resolve conflicts.