Learning Short-Term and Long-Term Facial Behaviors for Personality Traits Recognition

Personality is the integration of human psychological characteristics that influence a wide range of human behaviors. Previous automatic personality recognition methods either model personality only using frame-level or short segment-level facial behaviors, which ignores crucial long-term information, or extract long-term behaviors with a large number of frames being removed. In this paper, we first extract frame-wise facial attributes from a video. Then, we use statistic, spectral heatmaps and spectral vectors to encode video-level behaviors from the extracted frame-level facial attributes. Finally, the obtained video-level facial behavior representations are fed to Multilayer Perceptron (MLP) or Temporal Convolution Neural Network (TCN) for personality traits recognition. The experimental results on the ChaLearn First Impressions dataset show that the proposed approach which only used the facial attributes, has clear advantages over the static face image-based personality recognition approaches.