Welcome to the official repository of MAE-Face, a pre-trained model designed for generic facial representation.

MAE-Face leverages masked autoencoders, a self-supervised learning approach, on large-scale facial datasets to learn robust facial representation.

MAE-Face has achieved state-of-the-art performance for facial affective analysis across multiple datasets, including BP4D, DISFA, AffectNet, RAF-DB, Aff-Wild2, and Emo135, etc.

Moreover, its application is not restricted to facial affective analysis. Its versatility has the potential to extend to various facial tasks. It is welcome for you to explore the potential of MAE-Face and share your findings with us (through issues or contact us).

• Facial Action Unit Detection and Intensity Estimation from Self-supervised Representation
• A Unified Approach to Facial Affect Analysis: the MAE-Face Visual Representation

To use MAE-Face, clone this repository and install the required dependencies:

• PyTorch
• timm

To use the pre-trained MAE-Face model, download the .pth files from the Releases section. Use models_vit.py to create the model, and load the pre-trained weights from the .pth files. You can refer to example.py, which shows a simple example about how to load the pre-trained model.

For more details, please refer to the orignial MAE repository. For example, use main_finetune.py to fine-tune MAE-Face on your own dataset, or use main_pretrain.py for further pre-training.

If you use the above models, please cite the following papers:

@article{ma2022facial,
  title={Facial Action Unit Detection and Intensity Estimation from Self-supervised Representation},
  author={Ma, Bowen and An, Rudong and Zhang, Wei and Ding, Yu and Zhao, Zeng and Zhang, Rongsheng and Lv, Tangjie and Fan, Changjie and Hu, Zhipeng},
  journal={arXiv preprint arXiv:2210.15878},
  year={2022}
}

@inproceedings{ma2023unified,
  title={A Unified Approach to Facial Affect Analysis: The MAE-Face Visual Representation},
  author={Ma, Bowen and Zhang, Wei and Qiu, Feng and Ding, Yu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5923--5932},
  year={2023}
}

Bowen Ma