Facial Expression Recognition based on Landmarks

Abstract

Facial expression is presentation of people's emotion, which plays a vital role in person's daily communication. Therefore, facial expression recognition (FER) is becoming an increasingly significant task in contemporary society. Currently the majority of the proposed methods for facial expression recognition use deep convolutional neural networks (CNN) in a supervised learning fashion. In this paper, we try to answer two important questions. The first is that in human FER, what are the key features that mostly constitute an expression, if such features exist? The second question is to find a way to increase the FER accuracy. To this end, we propose a new FER framework that relies solely on facial landmarks. In this framework, in the training process, we first extract facial landmarks, and feed them into a shallow network for recognition/classification. We show that just by using 68 facial landmark points, it's possible to achieve state-of-the-art FER results, thus opening the possibility to further study human emotion cognition process. On the other hand, our framework also produce better results than typical deep CNN-based methods with fast implementation.