Self-supervised-motion

This is the repository that contains source code for the paper:

[AAAI 2024] Self-Supervised Bird’s Eye View Motion Prediction with Cross-Modality Signals

Installation

Create environment and install pytorch.

conda create --name sf python=3.8 -y
conda activate sf
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

Install spconv v1.2.1, pytorch3d and pypcd.

Install pointnet2

cd lib/pointnet2
python3 setup.py install

Other dependencies

pip install -r requirements.txt

Data Preparation

For detailed instructions refer to data_preparation.

Training

./train.sh

Citation

@article{fang2024self,
  title={Self-Supervised Bird's Eye View Motion Prediction with Cross-Modality Signals},
  author={Fang, Shaoheng and Liu, Zuhong and Wang, Mingyu and Xu, Chenxin and Zhong, Yiqi and Chen, Siheng},
  journal={arXiv preprint arXiv:2401.11499},
  year={2024}
}

Credits

This repository is based on the implementations of following repositories:

• motionnet
• flowstep3d
• pillar-motion