sit's Introduction

SIT

Paper

This is the official implementation of the paper: Social Interpretable Tree for Pedestrian Trajectory Prediction (AAAI 2022).

Quick Start

Requires:

Python== 3.6
numpy==1.16.4
torch==1.4.0

1) Install Packages

 pip install -r requirements.txt

2) Dataset

Please download the dataset and extract it into the directory './dataset/' like this:

./dataset/train/
./dataset/test/

Performance

Results on ETH-UCY and Stanford Drone Dataset:

	minADE	minFDE
ETH	0.39	0.62
HOTEL	0.14	0.22
UNIV	0.27	0.47
ZARA1	0.19	0.33
ZARA2	0.16	0.29
AVG ETH-UCY	0.23	0.38
SDD	9.13	15.42

SIT

Training & Evaluation

Suppose the training data is at ./dataset/. You can train and evaluate our model on the 'eth' dataset by below command:

bash train.sh 'eth'

Training on a single 2080Ti.

Acknowledgement

Thank for the pre-processed data provided by the works of PECNet.

Citation

If you find our work useful for your research, please consider citing the paper:

@inproceedings{sit,
  title={Social Interpretable Tree for Pedestrian Trajectory Prediction},
  author={Shi, Liushuai and Wang, Le and Long, Chengjiang and Zhou, Sanping and Zheng, Fang and Zheng, Nanning  and Hua, Gang},
  booktitle={Association for the Advance of Artificial Intelligence},
  year={2022}
}

sit's People

Contributors

Stargazers

Watchers

sit's Issues

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions

Hello and thank you for your inspiring work！

I encountered an issue in the dataset.py file, specifically in the following code block:

traj_new = np.array(trajs_new)
masks_new = np.array(masks_new)

The error message reads: "setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (18,) + inhomogeneous part."

I attempted to print trajs_new list and discovered that not all elements in the first dimension align with the expected batch size (512). There are elements greater than 512, some smaller than 512, and only a few that equal 512. Ideally, they should all be equal to 512, but I'm facing challenges in rectifying this issue.

Recommend Projects

lssiair / sit Goto Github PK