STGAT

STGAT: Modeling Spatial-Temporal Interactions for Human Trajectory Prediction

Correction

Our statement about Average Displacement Error (ADE) in the paper is wrong, and it should be RMSE or L2 distance (as in SocialAttention and SocialGan).

Requirements

• Python 3
• PyTorch (1.2)
• Matplotlib

Datasets

All the data comes from the SGAN model without any further processing.

How to Run

• First cd STGAT
• To train the model run python train.py (see the code to understand all the arguments that can be given to the command)
• To evalutae the model run python evaluate_model.py
• Using the default parameters in the code, you can get most of the numerical results presented in the paper. But a reasonable attention visualization may require trained for a longer time and tuned some parameters. For example, for the zara1 dataset and pred_len is 8 time-steps,, you can set num_epochs to 600 (line 36 in train.py), and the learning rate in step3 to 1e-4 (line 180 in train.py).
• The attachment folder contains the code that produces the attention figures presented in the paper
• Check out the issue of this repo to find out how to get better results on the ETH dataset.

Acknowledgments

All data and part of the code comes from the SGAN model. If you find this code useful in your research then please also cite their paper.

If you have any questions, please contact huangyingfan@outlook.com, and if you find this repository useful for your research, please cite the following paper:

@InProceedings{Huang_2019_ICCV,
author = {Huang, Yingfan and Bi, Huikun and Li, Zhaoxin and Mao, Tianlu and Wang, Zhaoqi},
title = {STGAT: Modeling Spatial-Temporal Interactions for Human Trajectory Prediction},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}