salesforce / woad-pytorch

This is the pytorch implementation of WOAD: Weakly Supervised Online Action Detection in Untrimmed Videos (CVPR2021).

WOAD: Weakly Supervised Online Action Detection in Untrimmed Videos

Introduction

• Pytorch implementation of WOAD: Weakly Supervised Online Action Detection in Untrimmed Videos

Environment

• The code is developed with CUDA V9.0, Python 3.6.3

Install

• pip install -r requirements.txt

Data Preparation

• Download Thumos14 annotations from here

• Download Thumos14reduced-I3D-JOINTFeatures from here

• Put the downloaded annotations under Thumos14reduced-Annotations/ and features under data/

Pretrained Models

• THUMOS'14 weakly-supervised model and strongly-supervised model

Evaluation

python eval.py --pretrained-ckpt MODEL_NAME

Training

python main.py --supervision SUPERVISION_TYPE --model-name NAME_TO_SAVE_MODEL

Citations

• If you find this codebase useful, please cite our paper:

@inproceedings{gao2021woad,
    title = {WOAD: Weakly Supervised Online Action Detection in Untrimmed Videos},
    author = {Mingfei Gao, Yingbo Zhou, Ran Xu, Richard Socher, Caiming Xiong},
    booktitle = {CVPR},
    year = {2021}
}

Contact

• Please send an email to [e-mail] if you have questions.

Acknowledgement

We referenced W-TALC for the code.