[Refactor] Camera keys

[filaPro]

1. Remove legacy keys from Collect3D: rect, Trv2c, P2.
2. Rename cam_intrinsic to cam2img to be consistent with depth2img and lidar2img. In next PRs this can unify image dependent functions like visualization and point fusion. More discussion in #791#issuecomment-887258262.

[codecov]

Codecov Report

Merging #805 (3d77518) into master (a876a47) will not change coverage.
The diff coverage is 54.54%.

@@           Coverage Diff           @@
##           master     #805   +/-   ##
=======================================
  Coverage   49.07%   49.07%           
=======================================
  Files         210      210           
  Lines       15961    15961           
  Branches     2547     2547           
=======================================
  Hits         7833     7833           
  Misses       7626     7626           
  Partials      502      502           

Flag	Coverage Δ
unittests	49.07% <54.54%> (ø)

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

Impacted Files	Coverage Δ
mmdet3d/apis/inference.py	43.21% <0.00%> (ø)
mmdet3d/datasets/nuscenes_mono_dataset.py	37.08% <ø> (ø)
mmdet3d/datasets/pipelines/formating.py	60.00% <ø> (ø)
mmdet3d/models/dense_heads/fcos_mono3d_head.py	12.06% <0.00%> (ø)
mmdet3d/models/detectors/single_stage_mono3d.py	16.66% <0.00%> (ø)
mmdet3d/core/visualizer/image_vis.py	71.83% <80.00%> (ø)
mmdet3d/datasets/pipelines/loading.py	90.61% <100.00%> (ø)
mmdet3d/datasets/pipelines/transforms_3d.py	90.41% <100.00%> (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 a876a47...3d77518. Read the comment docs.

[Wuziyi616]

I assume that you have used Ctrl+F to find and replace all the cam_intrinsic to cam2img? Also, please kindly test the code by browse_dataset.py on nuScene mono dataset and running the mono3d detection demo to check the visualization results. Thanks!

[filaPro]

I assume that you have used Ctrl+F to find and replace all the cam_intrinsic to cam2img?

Unfortunately, no, as we have some cam_intrinsic in KITTI, nuScenes and Lyft annotations. So I replaced them only after loading pipelines. So, after Collect3D we will not have cam_intrinsic. But for visualization and browse dataset will probably keep it.

Also, please kindly test the code by browse_dataset.py on nuScene mono dataset and running the mono3d detection demo to check the visualization results.

Ok will try to visualize.

[Wuziyi616]

I assume that you have used Ctrl+F to find and replace all the cam_intrinsic to cam2img?

Unfortunately, no, as we have some cam_intrinsic in KITTI, nuScenes and Lyft annotations. So I replaced them only after loading pipelines. So, after Collect3D we will not have cam_intrinsic. But for visualization and browse dataset will probably keep it.

Also, please kindly test the code by browse_dataset.py on nuScene mono dataset and running the mono3d detection demo to check the visualization results.

Ok will try to visualize.

That makes sense. Looking forward to the visualization results :)

[filaPro]

Hi @Tai-Wang, @Wuziyi616,

Monocular demo seems to be broken and not by my commits. Am I missing something?

[Tai-Wang]

Hi @filaPro , that's due to code refactoring for removing hacks in the previous codes. See #744 and #794 . The pretrained model link should be updated in the documentation (I will create a PR to fix that). You can try to pull the latest code and use the updated checkpoints. Everything seems fine from my side.

[filaPro]

You can try to pull the latest code and use the updated checkpoints

Really I pulled master and downloaded the checkpoint today. Will check one more time.

UPD. Sorry for that. Master branch is fine.

[Tai-Wang]

You can try to pull the latest code and use the updated checkpoints

Really I pulled master and downloaded the checkpoint today. Will check one more time.

UPD. Sorry for that. Master branch is fine.

Never mind :) Thank you for reminding me the link has not been updated. It really makes that example confusing.

[filaPro]

@Tai-Wang, looks like browse_dataset is also broken. Here we have not enough parameters for show_multi_modality_result call. Am I right?

As a temporal solution we can set bbox_mode='lidar' here. Do I need to make a separate PR?

BTW, have you checked this in current master? I'm getting invalid bbox rotations now. (But maybe I'm missing something again)

[Tai-Wang]

@filaPro That seems a legacy bug. It is not introduced by my refactoring, but I will create a PR to fix that.

After fixing the missing parameter, everything is ok for me. Have you re-generated the json files after merging the master? I have removed the hacks by adding some transformations in the pre-processing and post-processing. See more details in the compatibility documentation.

[filaPro]

@Wuziyi616, @Tai-Wang browse_dataset and mono_det_demo are fine.

[Wuziyi616]

The visualization seems to be good. Great job!