Home

Awesome

SGCN

Code for SGCN:Sparse Graph Convolution Network for Pedestrian Trajectory Prediction (CVPR2021)

Paper

Important Notice

The code will be pulled off shelves and upated to add more information, such as more detailed training process, a new proposed method for multimodal trajectory prediction, in our extensive journal version.

Environment

The code is trained and tested on RTX 2080Ti, Python 3.6.2, numpy 1.19.0, pytorch 1.2.0 and CUDA11.2.

SGCN

SGCN models the sparse graph in two aspects:

  1. a Spatial Sparse Graph to represent sparse interaction
  2. a Temprial Sparse Graph to represent diverse motion tendencies

framework

<div align='center'> <img src="images/sgcn.jpg"></img> </div> <br />

Sprarse Graph Learning

<div align='center'> <img src="images/sparse_graph_learning.jpg"></img> </div> <br />

Train

To train a model on a single dataset, such as eth, you can run below code in root directory of this project:

python train.py --dataset eth

or training on each dataset:

sh train.sh

Test

To evaluate your trained model or our pretrain model you can run:

python test.py

Note that the pretrain model lies in 'checkpoints'.

Download

You can download this project by :

git clone https://github.com/shuaishiliu/SGCN.git

Cite

If you find this repo useful, please consider citing our paper

@inproceedings{shi2021sgcn,
  title={Sparse Graph Convolution Network for Pedestrian Trajectory Prediction},
  author={Liushuai Shi and Le Wang and Chengjiang Long and Sanping Zhou and Mo Zhou and Zhenxing Niu and Gang Hua},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}