Pytorch implementation of the paper: "Self-Supervised Learning for Semi-Supervised Temporal Action Proposal" (CVPR-2021) [SSTAP-Paper]
- June 29, 2021： Slowfast101 feature used by the winners of the CVPR2020 ActivityNet Temporal Action Localization Challenge and the CVPR2021 ActivityNet Temporal Action Localization Challenge is [here]，and ActivityNet Challenge website is [here]. The features are not resized, the video frames are extracted at 15FPS, and the interval between each feature is 8 frames.
ActivityNet Challenge: [2021 champion solution--(PRN)], [2020 champion solution--(CBR-Net)]
Slowfast: [Slowfast Paper], [Slowfast Github]
The code runs correctly with:
- python 3.8.5
- pytorch 1.6.0
- torchvision 0.7.0
Other versions may also work.
Generate labeled/unlabeled data (you can also use our files directly)
python gen_unlabel_videos.py
bash SSTAP.sh | tee log_SSTAP.txt
BMN: Boundary-Matching Network
If our code is helpful for your reseach, please cite our paper:
@inproceedings{SSTAP,
title={Self-Supervised Learning for Semi-Supervised Temporal Action Proposal},
author={Wang, Xiang and Zhang, Shiwei and Qing, Zhiwu and Shao, Yuanjie and Gao, Changxin and Sang, Nong},
booktitle={CVPR},
year={2021}
}
@article{wang2020cbr,
title={CBR-Net: Cascade Boundary Refinement Network for Action Detection: Submission to ActivityNet Challenge 2020 (Task 1)},
author={Wang, Xiang and Ma, Baiteng and Qing, Zhiwu and Sang, Yongpeng and Gao, Changxin and Zhang, Shiwei and Sang, Nong},
journal={arXiv preprint arXiv:2006.07526},
year={2020}
}
@article{wang2021pro,
title={Proposal Relation Network for Temporal Action Detection},
author={Wang, Xiang and Qing, Zhiwu and Huang, Ziyuan and Feng, Yutong and Zhang, Shiwei and Jiang, Jianwen and Tang, Mingqian and Gao, Changxin and Sang, Nong},
journal={arXiv preprint arXiv:2106.11812},
year={2021}
}