Facial expression analysis using Decomposed Multiscale Spatiotemporal Networks

Wheidima Carneiro de Melo, Eric Granger, Miguel Bordallo López

Abstract: Highlights•A deep method is proposed with three variants (DMSN-A,B,C) to efficiently capture facial dynamics.•Using the variants, we build an architecture that can adapt to distinct facial behaviors.•We show that DMSN-C is effective for depression detection and DMSN-A is efficient for pain estimation.•Our architecture achieves similar results as the state-of-the-art at reduced computational costs.