Boosting facial action unit detection with CGAN-based data augmentation

Abstract

Contraction of the facial muscle movements is the key aspect of facial expression recognition tasks. The Facial Action Coding System (FACS) is the most widely used and accepted standard that provides a description of all major and minor visual changes in terms of action units (AUs) representing facial muscle movements. With the advancements of deep networks, the main problem shifted from detection or classification to finding sufficient amounts of data, especially when it comes to minor muscle movements on the face. This study employs Generative Adversarial Networks (GANs) as a data augmentation method for the task of AU detection on two spontaneous datasets (DISFA, BP4D) and one in-the-wild dataset (EmotioNet). Results show that AU detection scores increase using GANs when compared to using only traditional augmentation methods.