search

dvlab-research / UVTR

Unifying Voxel-based Representation with Transformer for 3D Object Detection (NeurIPS 2022)
227 stars 17 forks source link

mmdet3d version problem #8

Closed YongtaoGe closed 2 years ago

YongtaoGe commented 2 years ago

Hi, thanks for you work. I found this project is based on mmdet 0.17. I want to know if I use mmdet > 1.0, will I need to modify the code? mmdet3d > 1.0 seems have changed the coordinate system.

yanwei-li commented 2 years ago

Yes, you are right. It is suggested to use mmdet3d v0.17 for testing. If you use v1.x, due to the change in coordinate system, you need to change the code in 1. data info preparation, 2. nuscenes dataset, 3. box conversion in output.

YongtaoGe commented 2 years ago

Hi, thanks for your reply. It is appreciated if you could provide a code snippet for the box conversion in output. I found that it seems you already have trained your code in 1.0x in another issue.

yanwei-li commented 2 years ago

Sorry for the late reply. I have merged the modification to support mmdet3d v1.x. Please try it again~

YongtaoGe commented 2 years ago

@yanwei-li Thanks for your update. It really helps.

yanwei-li commented 2 years ago

Hi @YongtaoGe, I also update the data converter. You may need to regenerate the data info in mmdet v1.x following this 

python3 extra_tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes_unified
YongtaoGe commented 2 years ago

Thanks @yanwei-li. I will have a try.