Self-supervised Representation Learning from Videos for Facial Action Unit Detection, CVPR 2019 (oral)
We propose a Twin-Cycle Autoencoder (TCAE) that self-supervisedly learns two embeddings to encode the movements of AUs and head motions.
Given a source and target facial images, TCAE is tasked to change the AUs or head poses of the source frame to those of the target frame by predicting the AU-related and pose-related movements, respectively.
After training, the learned encoder can be used for AU detection. The extracted AU embedding from the encoder can be used for both AU detection and facial image retrieval.
If you use this code in your paper, please cite the following:
@inproceedings{li2019self,
title={Self-supervised Representation Learning from Videos for Facial Action Unit Detection},
author={Li, Yong and Zeng, Jiabei and Shan, Shiguang and Chen, Xilin},
booktitle={CVPR},
year={2019}
}