End-to-End Protocols and Performance Metrics For Unconstrained Face Recognition

Face recognition algorithms have received substantial attention over the past decade resulting in significant performance improvements. Arguably, improvement can be attributed to the wide spread availability of large face training sets, GPU computing to train state-of-the-art deep learning algorithms, and curation of challenging test sets that continue to push the state-of-the-art. Traditionally, protocol design and algorithm evaluation have primarily focused on measuring performance of specific stages of the biometric pipeline (e.g., face detection, feature extraction, or recognition) and do not capture errors that may propagate from face input to identification output in an end-to-end (E2E) manner. In this paper, we address this problem by expanding upon the novel open-set E2E identification protocols created for the IARPA Janus program. In particular, we describe in detail the joint detection, tracking, clustering, and recognition protocols, introduce novel E2E performance metrics, and provide rigorous evaluation using the IARPA Janus Benchmark C (IJB-C) and S (IJB-S) datasets.