Model weights for all trained methods (FastFlow3D and ZeroFlow) and their ablations are provided in their own GitHub repo.
To set up the file system, follow the Getting Started guide for BucketedSceneFlowEval. This will set up the Argoverse 2 and Waymo Open datasets in the correct format.
This project provides a docker image for training and evaluation using the Docerkfile in docker/Dockerfile [dockerhub]. The launch script ./launch.sh will start the docker container with the correct mounts and environment variables.
These mounts are
-v `pwd`:/project
runs pwd inside the script, getting the current directory, and ensures that it's mounted as /project inside the container.
The /efs/ mounts are for the Argoverse 2 and Waymo Open datasets. You must link Argoverse 2 so that inside the container it appears at /efs/argoverse2 and Waymo Open so that inside the container it appears at /efs/waymo_open_processed_flow. If you have these datasets in a different location, you can modify the /efs mount in launch.sh to point to the correct location.
The base system must have driver support for CUDA 11.3+ (you do not actually need CUDA installed on your base system, but you do need a driver version that supports the container install) and NVidia Docker installed.