Fully automatic 3D facial expression recognition using a region-based approach

Abstract

In this paper, we address the problem of automatic 3D facial expression recognition. Automatic 3D Facial Expression Recognition techniques are generally limited in that they require manual, precise landmark points. Here, we propose a framework capable of handling the potential imprecision of automatic landmarking techniques, thanks to a region approach. After an automatic feature point localization step, we cluster the face into several regions, chosen for their importance into the facial expression process, according to the Facial Action Coding System (FACS) and anatomic considerations. Then, we match those regions to reference models representing the six prototypical expressions using Iterative Closest Points (ICP). ICP tends to compensate the imprecisions in the face clustering relative to landmarks localization. Resulting matching scores are concatenated into a descriptor for the probe model. Finally, we use a standard classification tool; in our experiments, we used Support Vector Machines (SVM), and were able to provide comparable results to existing 3D FER methods over the same protocol, while being fully automatic.