Comparison of Face Recognition Loss Functions

Significant progresses have been made to face recognition algorithms in recent years. The progresses include the improvements of the solutions and the availability of more challenging databases. As the performance on previous benchmark databases, such as MPIE and LFW, saturates, more challenging databases are emerging and keep driving the advancements of face recognition solutions. The loss function considered in the solution usually plays the most critical role. We compare a few latest loss functions with the same feature embedding network by evaluating their performance on two recently-released databases, the IARPA Janus BenchmarkB (IJB-B) [1], and IARPA Janus BenchmarkC (IJB-C) [2], and highlight the directions for the next phase research.