Incremental Learning of People Identities

Abstract

Face recognition in unconstrained open-world settings is a challenging problem. Differently from the closed-set and open-set face recognition scenarios that assume that the face representations of known subjects have been manually enrolled in a gallery, the open-world scenario requires that the system learns identities incrementally from frame to frame, discriminate between known and unknown identities and automatically enrolls every new identity in the gallery, so to be able to recognize it every time it is observed again in the future. Performance scaling with large number of identities is likely to be needed in real situations. In this paper we discuss the problem and present a system that has been designed to perform effective open-world face recognition in real time at both small-moderate and large scale.