[Feature] Support ImVoxelNet on KITTI

[filaPro]

Reimplementation of ImVoxelNet for kitti-3d-car

[codecov]

Codecov Report

Merging #627 (9a40084) into master (c33d4ec) will increase coverage by 0.05%.
The diff coverage is 61.17%.

@@            Coverage Diff             @@
##           master     #627      +/-   ##
==========================================
+ Coverage   51.04%   51.10%   +0.05%     
==========================================
  Files         201      203       +2     
  Lines       15221    15307      +86     
  Branches     2468     2472       +4     
==========================================
+ Hits         7770     7822      +52     
- Misses       6933     6967      +34     
  Partials      518      518              

Flag	Coverage Δ
unittests	51.10% <61.17%> (+0.05%)	⬆️

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

Impacted Files	Coverage Δ
mmdet3d/models/detectors/imvoxelnet.py	56.36% <56.36%> (ø)
mmdet3d/models/necks/imvoxel_neck.py	66.66% <66.66%> (ø)
mmdet3d/models/detectors/__init__.py	100.00% <100.00%> (ø)
mmdet3d/models/necks/__init__.py	100.00% <100.00%> (ø)
mmdet3d/datasets/lyft_dataset.py	71.35% <0.00%> (-0.22%)	⬇️
mmdet3d/core/bbox/structures/utils.py	97.05% <0.00%> (ø)

---

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 c33d4ec...9a40084. Read the comment docs.

[Tai-Wang]

Overall looks good to me. Just need to fix some minor comments.
In addition, maybe we can inherit the kitti base dataset config in the imvoxelnet config?

[filaPro]

In addition, maybe we can inherit the kitti base dataset config in the imvoxelnet config?

I think there are too much point-cloud related transforms and we need only images.

Also, do I need to add a unit test?

[Tai-Wang]

In addition, maybe we can inherit the kitti base dataset config in the imvoxelnet config?

I think there are too much point-cloud related transforms and we need only images.

Also, do I need to add a unit test?

OK, then we can keep the current format for configs.
Unit tests are needed, we'd better also add them in this PR.
Besides, I observe that the linting failed. You can refer to this guidance for useful pre-commit tools to avoid this problem.

[Tai-Wang]

Almost ready to merge except consistent expression of Tensor in the docstring

[filaPro]

I will also run training on 8 GPUs one more time to be sure that refactoring Anchor3DGenerator and ConvModule not much influenced anything.

[filaPro]

I reverted anchor ranges, metrics are fine now. Can share model and log if needed.