Skip to content

Latest commit

 

History

History
26 lines (19 loc) · 1.89 KB

environment.md

File metadata and controls

26 lines (19 loc) · 1.89 KB
Raw

Environment Setup

Our implementation mainly relies on mmdetection3d==v0.17.1, the same as previous vision-based 3D detection methods. For detailed instructions, please refer to the official documentations of mmdetection3d.

However, we understand the pain for installation, because many mmlab modules need to be built from source and the version mismatches can cause confusing error messages.

  1. We recommend that you directly use our docker file.

If you are new to docker, I recommand the following bash scipts.

# Remember to update the DATA_ROOT and CKPTS_ROOT in Makefile to your own paths.
# Delete the AWS and WANDB options in Makefile if you don't use them.
make docker-build
make docker-dev
cd ../PF_Track
  1. If docker does not work for you, you may manually install the required packages referring to our docker file, then go to step 3. For your convenience, some important versions of our packages are:
  • Pytorch-related. pytorch==1.9.0, cuda==11.1, cudnn==8.
  • MMLab-related. (build from source recommended). MMCV==1.4.0, MMDetection==v2.24.1, MMSegmentation==v0.20.2, MMDetection3d==v0.17.1.
  1. Please also install these dependencies manally with pip.
  • nuScenes-related. nuscenes-devkit==1.1.7, motmetrics==1.1.3 (don't use higher versions of motmetrics, or it will cause nuscenes-devkit into bugs, which is a known nuScenes dependency issue.)
  • Visualization-tools. (optional). Install SimpleTrack locally via pip install -e ROOT_OF_SIMPLETRACK for BEV visualization.
  1. We thank DETR3D and PETR for open-sourcing. The projects/mmdet3d_plugin/ directory are the basic modules for 3D detection insipired by their implementations. We have already contained them for your convenience.