SE-STAD: A Simple and Efficient Pipeline to Build an End-to-End Spatial-Temporal Action Detector

Introduction

This repository hodes the official implementation of the paper "A Simple and Efficient Pipeline to Build an End-to-End Spatial-Temporal Action Detector".

We are continuing cleaning the code and we have released part of our code. Our paper is accepted to WACV 2023 and an arXiv version can be found at this link.

To-do list

• [ ] (Partially, in progress) Release the ssl part.
• [x] Release the baseline training code.

  Changelog

• 01/03/2022. Initial code release without SSL part.

Code Overview

Our code is based on MMAction2 and MMDetection2 with some major modification changes.

Installation

• Follow INSTALL.md for installing necessary dependencies and compiling the code.

To Reproduce Our Results on AVA

Download Features and Annotations

• Please follow the official repo from AVA Download Page and MMAction2 AVA prepare tutorial to prepare AVA dataset.

Training and Evaluation

• Train our SE-STAD with baseline part.

cd mmaction2
./run/train/slowfast_r50_fcos.sh

• Evaluate the trained model.

  cd mmaction2
  ./run/test/e2e_test.sh

• Train our SE-STAD with SSAD part.

  To be filled.

  Contact

  Chen-Lin Zhang (zclnjucs@gmail.com)

References

If you are using our code, please consider citing our paper.

@inproceedings{sui2023sestad,
  title={A Simple and Efficient Pipeline to Build an End-to-End Spatial-Temporal Action Detector},
  author={Sui, Lin, and Zhang, Chen-Lin and Gu, Lixin and Han Feng},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month={January},
  year={2023},
  pages={in press}
}