Open-set face identification with index-of-max hashing by learning

Large-scale face identification or 1-to-N matching where N is huge, plays a vital role in biometrics and surveillance. The system demands accurate and speedy matching where compact facial feature representation and a simple matcher are favored. On the other hand, most research considers closed-set identification that assumes that all identities of probe samples are enclosed in the gallery. On the contrary, open-set identification expects that some probe identities are not known to the system. This setup poses an additional challenge, where the system should be able to reject those probes that correspond to unknown identities. In this paper, we address the large-scale open-set face identification problem with a compact facial representation that is based on the index-of-maximum (IoM) hashing, which was designed for biometric template protection. To be specific, the existing random IoM hashing is advanced to a data-driven based hashing technique, where the hashed face code can be made compact and matching can be easily performed by the Hamming distance, which can offer highly efficient matching. Furthermore, since IoM hashing transforms the original facial features non-invertibly, the privacy of users can also be preserved. Along with IoM hashed face code, we explore several fusion strategies to address the open-set face identification problem. The comprehensive evaluations are carried out with three large-scale unconstrained face datasets, namely LFW, VGG2 and IJB-C.