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
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
| 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 |
| Method | Extra Data | Backbone | UCF101 | HMDB51 |
|---|---|---|---|---|
| VideoMAE | Kinetics-400 | ViT-S | 92.9 | 72.0 |
| VideoMAE | Kinetics-400 | ViT-B | 96.2 | 77.1 |
Please follow the instructions in INSTALL.md.
We provide pre-trained and fine-tuned models in MODEL_ZOO.md.
We provide the script for visualization in vis_kd.sh.
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}
}
This project is made available under the Apache 2.0 license.
This project is built upon VideoMAE. Thanks to the contributors of this great codebase.