FATAUVA-Net : An Integrated Deep Learning Framework for Facial Attribute Recognition, Action Unit Detection, and Valence-Arousal Estimation

Wei-Yi Chang, Shih-Huan Hsu, Jen-Hsien Chien; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 17-25

Facial expression recognition has been investigated for many years, and there are two popular models: Action Units (AUs) and the Valence-Arousal space (V-A space) that have been widely used. However, most of the databases for estimating V-A intensity are captured in laboratory settings, and the benchmarks "in-the-wild" do not exist. Thus, the First Affect-In-The-Wild Challenge released a database for V-A estimation while the videos were captured in wild condition. In this paper, we propose an integrated deep learning framework for facial attribute recognition, AU detection, and V-A estimation. The key idea is to apply AUs to estimate the V-A intensity since both AUs and V-A space could be utilized to recognize some emotion categories. Besides, the AU detector is trained based on the convolutional neural network (CNN) for facial attribute recognition. In experiments, we will show the results of the above three tasks to verify the performances of our proposed network framework.