Flow4D: Leveraging 4D Voxel Network for LiDAR Scene Flow Estimation (RA-L 2025)

This repository contains the code for the Flow4D paper (RA-L 2025)

Notice

Flow4D has been integrated into OpenSceneFlow. Please visit the OpenSceneFlow repository for the latest updates and developments.

This repo saved README, and quick core file in Flow4D for a quick reference. The old source code branch is also available here.

Requirements

This code is based on DeFlow.
Please follow the installation instructions from the DeFlow repository.

Additionally, you need to install spconv 2.3.6.
You can find the installation instructions here: spconv.

Training

To train the model, use the following command:

python train.py model=flow4d lr=1e-3 epochs=15 batch_size=8 num_frames=5 loss_fn=deflowLoss "voxel_size=[0.2, 0.2, 0.2]" "point_cloud_range=[-51.2, -51.2, -3.2, 51.2, 51.2, 3.2]"

Inference

To perform inference, use the following command:

python eval.py checkpoint=path_to_checkpoint av2_mode=(val, test)

Replace path_to_checkpoint with the actual path to your checkpoint file and choose either val or test.

Gratitude

This code is based on the DeFlow code by Qingwen Zhang. We extend our deepest gratitude to her.
Additionally, we would like to express our sincere thanks to Kyle Vedder et al. for hosting and providing extensive support for Argoverse2 2024 Scene Flow Challenge