Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations
Official implementation for our CVPR2024 paper: "Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations". [Arxiv]
- Python 3.9
- Pytorch >= 2.0
# git clone this repository
git clone https://github.com/kwwcv/SelfMotion
cd SelfMotion
- Download the nuScenes data.
# modified the following paths in gen_data.py, gen_GSdata.py, and data_utils.py
# sys.path.append('root_path/SelfMotion')
# sys.path.append('root_path/SelfMotion/nuscenes-devkit/python-sdk/')
-
Run command
python data/gen_data.pyto generate preprocessed BEV data for validating, and testing. Refer to MotionNet andpython data/gen_data.py -hfor detailed instructions. -
Install the ground segmentation algorithm following Patchwork++. One can also try removing the ground points by simply setting a threshold along the Z-axis.
# modified the following path in gen_GSdata.py
# patchwork_module_path = "root_path/patchwork-plusplus/build/python_wrapper"
- Run command
python data/gen_GSdata.pyto generate preprocessed ground-removed BEV data for training.
python train.py --train_data [ground removal bev training folder] --test_data [bev validation folder] \
--log --log_path [path to save log] --if_cluster --if_forward --if_reverse
python test.py --data [bev testing folder] --model [model path] \
--log_path [path to save results]
@misc{wang2024selfsupervised,
title={Self-Supervised Class-Agnostic Motion Prediction with Spatial and Temporal Consistency Regularizations},
author={Kewei Wang and Yizheng Wu and Jun Cen and Zhiyu Pan and Xingyi Li and Zhe Wang and Zhiguo Cao and Guosheng Lin},
year={2024},
eprint={2403.13261},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Our project is based on MotionNet
The optimal transport solver is adopted from Self-Point-Flow
This project is licensed under NTU S-Lab License 1.0
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent