[Enhance] Support rotate points without bbox in GlobalRotScaleTrans

[Wuziyi616]

In the current implementation of GlobalRotScaleTrans, we will rotate points only when bbox is presented in input_dict (here). However, in PC Seg, we need to rotate points without bbox. In this PR, I add a if condition for this situation. I also add some input check & perform params transformation in __init__ to save future time.

[Wuziyi616]

I discover that in this unit test, someone has tested GlobalRotScaleTrans and the expected_points don't perform rotation. So, I wonder if don't rotate points when bbox is not presented is the desired behavior of this function?

[codecov]

Codecov Report

Merging #540 (85dfa52) into master (43d7953) will increase coverage by 0.09%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #540      +/-   ##
==========================================
+ Coverage   50.79%   50.89%   +0.09%     
==========================================
  Files         192      192              
  Lines       14601    14609       +8     
  Branches     2378     2379       +1     
==========================================
+ Hits         7417     7435      +18     
+ Misses       6691     6683       -8     
+ Partials      493      491       -2     

Flag	Coverage Δ
unittests	50.89% <100.00%> (+0.09%)	⬆️

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

Impacted Files	Coverage Δ
mmdet3d/core/points/base_points.py	82.38% <100.00%> (+0.10%)	⬆️
mmdet3d/datasets/pipelines/transforms_3d.py	89.89% <100.00%> (+2.19%)	⬆️

---

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 43d7953...85dfa52. Read the comment docs.

[Tai-Wang]

I discover that in this unit test, someone has tested GlobalRotScaleTrans and the expected_points don't perform rotation. So, I wonder if don't rotate points when bbox is not presented is the desired behavior of this function?

I guess it should be a legacy issue. We don't have a separate class for points before, so previously some transformations related to points are achieved with bounding boxes (a hack workaround). Similar problems exist here. Maybe we can gradually refactor these similar operations in the future.