A Dynamic Approach for Detecting Naturalistic Affective States from Facial Videos during HCI

Significant progress has been made in automatic facial expression analysis using facial images and videos. The recognition reliability of most current approaches is still poor in naturalistic expressions compared to acted ones. Most of these methods use a static image of each expression that captures the characteristic image at the apex. However, according to psychologists, analyzing a sequence of images in a dynamic manner produces more accurate and robust recognition of facial affect expressions. In this paper, a new dynamic model is proposed for detecting naturalistic affect expressions. The Local Binary Pattern in Three Orthogonal Planes (LBP-TOP) is considered for modeling appearance and motion of facial features. The International Affective Picture System (IAPS) collection was used as stimulus for triggering naturalistic affective states. The dynamic approach produced an improvement of 16% for valence classification and 22% for arousal classification over previous studies.