Giters

hova88 / PointPillars_MultiHead_40FPS

A REAL-TIME 3D detection network [Pointpillars] compiled by CUDA/TensorRT/C++.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Train a multihead model using the KITTI dataset

cyhasuka opened this issue · comments

commented

Hello,
Thank you for the great work you've done!

I'm a beginner, and I currently have a model that I trained by myself based on the KITTI dataset using the methods provided by the OpenPCDet project team. However, it doesn't work properly with your code.

I tried to train a multi-head model using the KITTI dataset by modifying the pointpillar.yaml file with DENSE_HEAD: NAME: AnchorHeadMulti and added RPN_HEAD_CFGS based on the three object categories in the KITTI dataset. The model trained successfully and was converted into a TensorRT engine, but it doesn't run correctly.

The output time, IDs and scores are confusing, and the data in the demo_boxes.txt output file are all zeros.Here is an example output:

------------------------------------
Module        Time        
------------------------------------
Preprocess    8.81295  ms
Pfe           534.183  ms
Scatter       3.85605  ms
Backbone      80.4403  ms
Postprocess   8.5541   ms
Summary       636.175  ms
------------------------------------
id: 0 labels: 2
      scores: 0.071332
id: 1 labels: 2
      scores: 0.0552087
id: 2 labels: 0
      scores: 0.117978
id: 3 labels: 0
      scores: 0.111822
id: 4 labels: 0
      scores: 0.110663
id: 5 labels: 0
      scores: 0.108945
.......
id: 2476 labels: 0
      scores: 0.0462514
id: 2504 labels: 539768118
      scores: 0.161591
id: 2516 labels: 1685202208
      scores: 0.161591
id: 2528 labels: 16873488
      scores: 0.161591
id: 2544 labels: 17134530
      scores: 0.646273
id: 3157 labels: 6146
      scores: 0.63662
id: 3177 labels: 255
      scores: 0.5
id: 3229 labels: 7
      scores: 0.5
id: 3233 labels: 2
      scores: 0.5
id: 3529 labels: -128
      scores: 0.166667
id: 3541 labels: 6
      scores: 0.0416667
id: 3549 labels: 0
      scores: 1

demo_boxes.txt:

0 0 0 0 0 0 0 
0 0 0 0 0 0 0 
......
0 0 0 0 0 0 0

Could you please advise on how to resolve these issues? Any guidance you can provide would be greatly appreciated.

Thank you for your help!

commented

checkout the RPN_HEAD_CFGS classes layout in cbgs_pp_multihead.yaml and the output feature layout from backbone in postprocess.cu. They are all one-to-one correspondence, like

  • cls_pred_0 := ['car']
  • cls_pred_12 := ['truck', 'construction_vehicle']

If you want to modify the output categories from nuscenes to kitti, the following points need to be noted.

In addition, I don't know your full running situation, maybe there also have some other problems. Please be patient with this, it's may a great learning process for you~

Good luck.

commented

Thank you very much for your suggestions and proposed changes.
I will follow your approach and make relevant attempts, and will provide updates on the progress below. Once again, thank you for your generous assistance!