-
Notifications
You must be signed in to change notification settings - Fork 1.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Try to train CenterPoint on kitti #486
Comments
base = [ point_cloud_range = [0, -39.68, -3, 69.12, 39.68, 1] model = dict( train_pipeline = [ data = dict( lr = 0.001 |
'centerpoint_02pillar_second_secfpn_kitti.py' in my config file is modified from centerpoint_02pillar_second_secfpn_nus.py,only changed some params (such as class names) to adapt KITTI, and I didn't change the architecture. |
Hi, we are also working on training CenterPoint on KITTI. _2021-04-27 13:54:02,527 - mmdet - INFO - Start running, host: root@b181bde9e289, work_dir: /content/gdrive/My Drive/Colab_Notebooks/mmdet3d-colab/mmdetection3d/work_dirs/centerpoint_nusc2kitti_02pillar |
The voxel size in my config file is [0.16, 0.16, 4], point cloud range is [0, -40.96, -1, 81.92, 40.96, 3], Pillar feature net output shape and grid size is [512, 512]. I'm sorry that I didn't meet this error, may be caused by pytorch version according to https://blog.csdn.net/yyhaohaoxuexi/article/details/107460836 @ZhangYu1ing |
One implementation here https://github.com/tianweiy/CenterPoint-KITTI I played with mmdet3d in the past but I don't get a clean codebase to release at the moment. Additionally, PointPillars just doesn't work too well on KITTI so you may consider playing with the VoxelNet backbone. |
That means that the hard_voxelize code is not compiled correctly for your GPU. |
Thanks very much for reply and sharing the code ! @tianweiy Another question is what is the mAP difference roughly between pillars and voxelnet on KITTI according to your experiment result? |
I only played with PointPillars in the Det3D repo. I am not able to get upper than 75 mAP. I will add model zoo to the repo soon. |
@guojingming @Tai-Wang Thanks for giving the advice. We noticed that the problem only be raised up in Google Colab. But it could work on PC. |
Hi, I am trying to implement your provided code. I created the KITTI data following the instruction in OpenPCDet. However, there is a bug when I run the tools/train.py, which is shown below. Do you have any ideas? Thanks a lot. 2021-04-28_ 15:44:55,801 INFO Start training tools/cfgs/kitti_models/centerpoint(default) |
Maybe, there are some version compatibility issue. I will check today |
No worries, the problem has been solved. The problem is related to the path of Kitti Dataset. I modified the DATA_PATH in tools/cfgs/dataset_configs/kitti_dataset.yaml This is my attempt to solve this problem. I am not sure if there are another issues also lead to this problem? |
I tried to train CenterPoint on kitti, when finished training, the val result is far from expected.
Use PillarFeatureNet as feature extracter, second as backbone and second_fpn as neck, train on 3-class kitti, cyc lr and 80 epochs, eval by strict 3D iou.
Got mAP of Car is 69.17, Ped is 30.68, Cyc is 53.53).
Are there problems with my config file?
The text was updated successfully, but these errors were encountered: