STAR

Code for Spatio-Temporal Graph Transformer Networks for Pedestrian Trajectory Prediction

Environment

The code is tested on GTX 1080Ti, Python 3.6.3, numpy 1.17.5, pytorch 1.1.0 and CUDA9.0.

Train

The Default settings are to train on ETH-univ dataset.

Data cache and models will be stored in the subdirectory "./output/eth/" by default. Notice that for this repo, we only provide implementation on GPU.

git clone https://github.com/Majiker/STAR.git
cd STAR
python trainval.py

Configuration files are also created after the first run, arguments could be modified through configuration files or command line. Priority: command line > configuration files > default values in script.

The datasets are selected on arguments '--test_set'. Five datasets in ETH/UCY including [eth, hotel, zara1, zara2, univ].

Example

This command is to train model for ETH-hotel and start test at epoch 10. For different dataset, change 'hotel' to other datasets named in the last section.

python trainval.py --test_set hotel --start_test 10

During training, the model for Best FDE on the corresponding test dataset would be record.

Cite STAR

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

@inproceedings{
    YuMa2020Spatio,
    title={Spatio-Temporal Graph Transformer Networks for Pedestrian Trajectory Prediction},
    author={Cunjun Yu and Xiao Ma and Jiawei Ren and Haiyu Zhao and Shuai Yi},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    month = {August},
    year={2020}
}

Reference

The code base heavily borrows from SR-LSTM

GIMPS / STAR