PETR icon indicating copy to clipboard operation
PETR copied to clipboard

Published 20 hours ago • verified publisher icon megvii-research

Reproduce PETRv1 result using refactored MMDetection3D codebase

Open WUMINGCHAzero opened this issue 2 years ago • 4 comments

Thanks for the great work! I tried to use MMDetection3D v1.0.0rc5 to train and evaluate PETRv1 using config:petr/petr_r50dcn_gridmask_p4.py, however, I only got mAP of 0.3011(v.s. 0.3174 in config) and NDS of 0.3488(v.s. 0.3665 in config). I used nuscenes_infos_train.pkl and nuscene_infos_val.pkl generated by MMDetection3D v1.0.0rc5 and did not do any modification to the code and config (model was trained on 8 V100 GPU). So, is there anything I need to change when training PETRv1 and PETRv2 on MMDetection3D v1.0.0rc5? I would like to use the latest version of MMDetection3D instead of the old one due to its advanced features.

Many thanks!

WUMINGCHAzero avatar Nov 07 '22 06:11 WUMINGCHAzero

Hi, Please check my investigation in https://github.com/WangYueFt/detr3d/issues/31. It should help!

yihongXU avatar Nov 10 '22 09:11 yihongXU

Hi, When use the MMDetection3D v1.0.0rc5, the reverse_angle must be False. https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_r50dcn_gridmask_p4.py#L139

yingfei1016 avatar Nov 17 '22 07:11 yingfei1016

Hi, When use the MMDetection3D v1.0.0rc5, the reverse_angle must be False. https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_r50dcn_gridmask_p4.py#L139

Can you please explain why?

tyxsspa avatar Dec 01 '22 07:12 tyxsspa

Hi, When use the MMDetection3D v1.0.0rc5, the reverse_angle must be False. https://github.com/megvii-research/PETR/blob/main/projects/configs/petr/petr_r50dcn_gridmask_p4.py#L139

Can you please explain why?

mmdet3d reconstructed the coordinate system at v1.0.0rc5, and the definition of angle has changed. You can refer to https://zhuanlan.zhihu.com/p/491614921

yingfei1016 avatar Dec 01 '22 08:12 yingfei1016