Official PyTorch Implementation of Mask Again: Masked Knowledge Distillation for Masked Video Modeling (ACM MM 2023).

Mask Again: Masked Knowledge Distillation for Masked Video Modeling [PDF]
Xiaojie Li^1,2, Shaowei He^1, Jianlong Wu^1*, Yue Yu^2, Liqiang Nie^1*, Min Zhang^1
^1Harbin Institute of Technology, Shenzhen, ^2Peng Cheng Laboratory *Corresponding Author

🚀 Main Results

✨ Kinetics-400

Method	Extra Data	Backbone	Resolution	#Frames x Clips x Crops	Top-1	Top-5
VideoMAE	no	ViT-S	224x224	16x5x3	78.7	93.6
VideoMAE	no	ViT-B	224x224	16x5x3	81.0	94.6

✨ UCF101 & HMDB51

Method	Extra Data	Backbone	UCF101	HMDB51
VideoMAE	Kinetics-400	ViT-S	92.9	72.0
VideoMAE	Kinetics-400	ViT-B	96.2	77.1

🔨 Installation

Please follow the instructions in INSTALL.md.

📍Model Zoo

We provide pre-trained and fine-tuned models in MODEL_ZOO.md.

👀 Visualization

We provide the script for visualization in vis_kd.sh.

✏️ Citation

If you find this project useful for your research, please considering leaving a star⭐️ and citing our paper:

@inproceedings{li2023mask,
  title={Mask Again: Masked Knowledge Distillation for Masked Video Modeling},
  author={Li, Xiaojie and He, Shaowei and Wu, Jianlong and Yu, Yue and Nie, Liqiang and Zhang, Min},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={2221--2232},
  year={2023}
}

🔒 License

This project is made available under the Apache 2.0 license.

👍 Acknowledgements

This project is built upon VideoMAE. Thanks to the contributors of this great codebase.

xiaojieli0903/MaskAgain