On Fisher vector encoding of binary features for video face recognition

摘要

Several approaches have been proposed for face recognition in videos. Fisher vector (FV) encoding of local Scale Invariant Feature Transforms (SIFT) is among the best performing ones. Aiming at speed up the computation time of this approach, a method based on FV encoding of binary features was recently introduced. By using Binary Robust Independent Elementary Features (BRIEF), this method gained in efficiency but lost in accuracy. FV representation of binary features demands appropriated mathematical tools, which are not as easy available as for continuous features. This paper introduces a new way for obtaining FV encoding of binary features that is still efficient and also accurate. We show that BRIEF combined with FV are discriminative enough, and provide as good performance as the one obtained by using SIFT features for video face recognition. Besides, we discuss several insights and promising lines of future work in regard to FV encoding of binary features.