Facial expression recognition using temporal POEM features

In this work, we propose a novel approach for Facial Expression Recognition (FER). We introduce the TPOEM (Temporal Patterns of Oriented Edge Magnitudes) features, which are volumetric extensions of the known POEM features by exploring adjacent frames and also temporal derivatives. To cope with the increase in the code length produced by TPOEM, we also present a novel coding scheme for the binary codes. TPOEM features are computed within non-overlapping image patches, and the final classification is achieved by combining the per-patch scores. The results showed increased accuracy compared to or better than state-of-the-art methods, while keeping the execution time low. (C) 2017 Elsevier B.V. All rights reserved.