ExpressionFlow: A Microexpression Descriptor for Efficient Recognition

Microexpressions are involuntary facial movements that often reflect a person’s true emotions. Their fleeting nature and subtle shifts, however, make them challenging to detect. Our earlier work, the Facial Dynamics Map, represented a microexpression by estimating dense optical flow. Although it achieved high prediction accuracy, it was inefficient in feature extraction and lacked magnitude information. In this chapter, we address these issues by proposing ExpressionFlow, a novel descriptor which directly captures the dominant motion patterns in microexpression image sequences. Geometrically intuitive and relatively easy to implement, ExpressionFlow reflects the nature of microexpressions while preserving complete information. Comparative experiments on four benchmark datasets suggest that our method attains the best performance in real-time compared with other state-of-the-art algorithms.