Unsupervised Pre-training for Temporal Action Localization (UP-TAL)

PyTorch Implementation of paper:

Unsupervised Pre-training for Temporal Action Localization Tasks (CVPR2022)

Can Zhang, Tianyu Yang, Junwu Weng, Meng Cao, Jue Wang and Yuexian Zou*.

Updates

• We will release our codes and models soon, please stay tuned!

Highlights

• This is the FIRST work focusing on Unsupervised Pre-training for Temporal Action Localization (UP-TAL).

• We design an intuitive and effective pretext task customized for TAL, called Pseudo Action Localization (PAL).

• Our PAL features transfer well on various TAL tasks: Temporal Action Detection (TAD), Action Proposal Generation (APG) and Video Grounding (VG).

TLDR

Given a video (), we randomly sample two pseudo action regions from it and then paste them onto another two pseudo background videos ( & ) at various temporal locations and scales. PAL learns temporal equivariant features by aligning pseudo action region features ( & ) and maximizing the agreement between region features of the same video but with different backgrounds.

Other Info

Citation

Please [★star] this repo and [cite] the following paper if you feel our PAL useful to your research:

@inproceedings{zhang2022pal,
    title     = {Unsupervised Pre-training for Temporal Action Localization Tasks},
    author    = {Zhang, Can and Yang, Tianyu and Weng, Junwu and Cao, Meng and Wang, Jue and Zou, Yuexian},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2022}
}

Contact

For any questions, please feel free to open an issue or contact:

Can Zhang: zhang.can.pku@gmail.com