Recursive binary template embedding for face image sets

In this paper, we focus on the task of generating compact binary templates for face image sets. Existing set-based templates are extracted from a high-dimensional real-value feature space, and sometimes with complex structure. The process is not very efficient for large-scale applications. Existing binary templates are generated via either a sequence or a tree of projections. The recursive tree structure achieves higher recognition performance, but the number of projections increases exponentially with the binary code length. Thus, we propose a recursive binary embedding algorithm exhibiting an increased recognition power while restricting the number of projections to increase linearly with the code length. Moreover, the proposed embedding can be easily plugged into the modern deep learning architectures for set-based face recognition. Experimental results using the challenge IJB-A dataset illustrate the effectiveness and efficiency of our proposed template for face verification, closed set identification, and open set identification.