VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
Zhan Tong, Yibing Song, Jue Wang, Limin Wang
Nanjing University, Tencent AI Lab
[2023.1.16] Code and pre-trained models for Action Detection are available!
[2022.11.20] 👀 VideoMAE is integrated into 
[2022.10.25] 👀 VideoMAE is integrated into MMAction2, the results on Kinetics-400 can be reproduced successfully.
[2022.10.20] The pre-trained models and scripts of ViT-S and ViT-H are available!
[2022.10.19] The pre-trained models and scripts on UCF101 are available!
[2022.9.15] VideoMAE is accepted by NeurIPS 2022 as a spotlight presentation! 🎉
[2022.8.8] 👀 VideoMAE is integrated into official 🤗HuggingFace Transformers now!
[2022.7.7] We have updated new results on downstream AVA 2.2 benchmark. Please refer to our paper for details.
[2022.4.24] Code and pre-trained models are available now!
[2022.3.24] Code and pre-trained models will be released here. Welcome to watch this repository for the latest updates.
VideoMAE performs the task of masked video modeling for video pre-training. We propose the extremely high masking ratio (90%-95%) and tube masking strategy to create a challenging task for self-supervised video pre-training.
VideoMAE uses the simple masked autoencoder and plain ViT backbone to perform video self-supervised learning. Due to the extremely high masking ratio, the pre-training time of VideoMAE is much shorter than contrastive learning methods (3.2x speedup). VideoMAE can serve as a simple but strong baseline for future research in self-supervised video pre-training.
VideoMAE works well for video datasets of different scales and can achieve 87.4% on Kinects-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51. To our best knowledge, VideoMAE is the first to achieve the state-of-the-art performance on these four popular benchmarks with the vanilla ViT backbones while doesn't need any extra data or pre-trained models.
| Method | Extra Data | Backbone | Resolution | #Frames x Clips x Crops | Top-1 | Top-5 |
|---|---|---|---|---|---|---|
| VideoMAE | no | ViT-S | 224x224 | 16x2x3 | 66.8 | 90.3 |
| VideoMAE | no | ViT-B | 224x224 | 16x2x3 | 70.8 | 92.4 |
| VideoMAE | no | ViT-L | 224x224 | 16x2x3 | 74.3 | 94.6 |
| VideoMAE | no | ViT-L | 224x224 | 32x1x3 | 75.4 | 95.2 |
| Method | Extra Data | Backbone | Resolution | #Frames x Clips x Crops | Top-1 | Top-5 |
|---|---|---|---|---|---|---|
| VideoMAE | no | ViT-S | 224x224 | 16x5x3 | 79.0 | 93.8 |
| VideoMAE | no | ViT-B | 224x224 | 16x5x3 | 81.5 | 95.1 |
| VideoMAE | no | ViT-L | 224x224 | 16x5x3 | 85.2 | 96.8 |
| VideoMAE | no | ViT-H | 224x224 | 16x5x3 | 86.6 | 97.1 |
| VideoMAE | no | ViT-L | 320x320 | 32x4x3 | 86.1 | 97.3 |
| VideoMAE | no | ViT-H | 320x320 | 32x4x3 | 87.4 | 97.6 |
Please check the code and checkpoints in VideoMAE-Action-Detection.
| Method | Extra Data | Extra Label | Backbone | #Frame x Sample Rate | mAP |
|---|---|---|---|---|---|
| VideoMAE | Kinetics-400 | ✗ | ViT-S | 16x4 | 22.5 |
| VideoMAE | Kinetics-400 | ✓ | ViT-S | 16x4 | 28.4 |
| VideoMAE | Kinetics-400 | ✗ | ViT-B | 16x4 | 26.7 |
| VideoMAE | Kinetics-400 | ✓ | ViT-B | 16x4 | 31.8 |
| VideoMAE | Kinetics-400 | ✗ | ViT-L | 16x4 | 34.3 |
| VideoMAE | Kinetics-400 | ✓ | ViT-L | 16x4 | 37.0 |
| VideoMAE | Kinetics-400 | ✗ | ViT-H | 16x4 | 36.5 |
| VideoMAE | Kinetics-400 | ✓ | ViT-H | 16x4 | 39.5 |
| VideoMAE | Kinetics-700 | ✗ | ViT-L | 16x4 | 36.1 |
| VideoMAE | Kinetics-700 | ✓ | ViT-L | 16x4 | 39.3 |
| Method | Extra Data | Backbone | UCF101 | HMDB51 |
|---|---|---|---|---|
| VideoMAE | no | ViT-B | 91.3 | 62.6 |
| VideoMAE | Kinetics-400 | ViT-B | 96.1 | 73.3 |
Please follow the instructions in INSTALL.md.
Please follow the instructions in DATASET.md for data preparation.
The pre-training instruction is in PRETRAIN.md.
The fine-tuning instruction is in FINETUNE.md.
We provide pre-trained and fine-tuned models in MODEL_ZOO.md.
We provide the script for visualization in vis.sh. Colab notebook for better visualization is coming soon.
Zhan Tong: [email protected]
Thanks to Ziteng Gao, Lei Chen, Chongjian Ge, and Zhiyu Zhao for their kind support.
This project is built upon MAE-pytorch and BEiT. Thanks to the contributors of these great codebases.
The majority of this project is released under the CC-BY-NC 4.0 license as found in the LICENSE file. Portions of the project are available under separate license terms: SlowFast and pytorch-image-models are licensed under the Apache 2.0 license. BEiT is licensed under the MIT license.
If you think this project is helpful, please feel free to leave a star⭐️ and cite our paper:
@inproceedings{tong2022videomae,
title={Video{MAE}: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training},
author={Zhan Tong and Yibing Song and Jue Wang and Limin Wang},
booktitle={Advances in Neural Information Processing Systems},
year={2022}
}
@article{videomae,
title={VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training},
author={Tong, Zhan and Song, Yibing and Wang, Jue and Wang, Limin},
journal={arXiv preprint arXiv:2203.12602},
year={2022}
}