FedRSU: Federated Learning for Scene Flow Estimation on Roadside Units

This is the official repository for our work "[FedRSU: Federated Learning for Scene Flow Estimation on Roadside Units]"(https://arxiv.org/abs/2401.12862)

FedRSU has been accepted to "IEEE Transactions on Intelligent Transportation Systems (T-ITS)"!

Installation

conda create --name fedrsu python=3.7 cmake=3.22.1

# torch
conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge

# pypcd
git clone https://github.com/klintan/pypcd.git
cd pypcd
pip install python-lzf
python setup.py install

# pointnet2
cd lib/pointnet2
python3 setup.py install
cd ../../

# others
opencv-python-headless, open3d, matplotlib, tensorboardx, pyyaml, pyquaternion

Datasets

Our dataset is now available through:

• Google Drive
• Baidu Netdisk (提取码: 56js)

Download and put the unzipped four directories in ./data.

The dataset downloaded here is a zip file composed of four separate directories: Dair-V2X, LUMPI, IPS300+ and Campus. Due to acknowledgement reasons, we haven't include IPS300+ in this initial version. We will update it in the next version soon.

Training

conda activate fedrsu

For non-distibuted setting: (Recommended)

> single gpu
python train.py --config ./config/example.yaml --alg fedavg --gpu 0

For distributed / Multi-GPU setting:

> multiple gpu
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 train.py --config ./example_ddp.yaml --ddp --alg fedavg --gpu 0

Citation

@article{fang2024fedrsu,
  title={FedRSU: Federated Learning for Scene Flow Estimation on Roadside Units},
  author={Fang, Shaoheng and Ye, Rui and Wang, Wenhao and Liu, Zuhong and Wang, Yuxiao and Wang, Yafei and Chen, Siheng and Wang, Yanfeng},
  journal={arXiv preprint arXiv:2401.12862},
  year={2024}
}