Face recognition performance has recently demonstrated parity with human performance for face verification on the well-established Labeled Faces in the Wild (LFW) dataset. However, this dataset (and others) have been constructed by pre-filtering imagery and video using an off-the-shelf face detectors to provide initial annotations, which bias these datasets to include only near-frontal faces. Operational face recognition datasets (for example, those encountered in video surveillance) include a large amount of diversity in pose, occlusion and image artifacts that limit the detection rate for classic face detectors such as Viola-Jones or PittPatt. Face recognition performance on these “media in the wild” datasets are significantly worse than the academic counterparts. The IARPA Janus program was created to address this performance gap, and we continue to extend the state of the art in this challenging research problem.