MambaFlow: A Novel and Flow-guided State Space Model for Scene Flow Estimation

Notice

This repository has been merged into a new repository. For the latest features, updates, and developments, please visit OpenSceneFlow.

All future development work will be conducted in the new repository.

Requirements

This code is based on Flow4D.
Please follow the installation instructions from the Flow4D repository.

Additionally, you need to install mamba-ssm 2.2.2 and causal-conv1d 1.4.0.
You can find the installation instructions here: mamba.

Training

To train the model, use the following command:

python 1_train_mambaflow.py

Inference

To perform inference, use the following command:

python 2_eval_mambaflow.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.

Visualization

To visualize scene flow:

1. Configure conf/save.yaml with model name, dataset path and weights path and use the following command to generate inference result:

   pathon save.py

2. Running following command to visualize scene flow:

   python tools/visualization.py --res_name 'your-method' --data_dir /path/to/dataset

• --res_name should match the name in save.yaml
• Ground truth will be generated if --res_name is not specified or set as flow

Gratitude

This code is based on the DeFlow code by Qingwen Zhang and the Flow4D code by dgist-cvlab. We extend our sincere appreciation to both teams for their contributions. Special thanks to Kyle Vedder et al. for organizing and providing support throughout the Argoverse2 2024 Scene Flow Challenge.