The availability of large annotated datasets and affordable computation power\nhave led to impressive improvements in the performance of CNNs on various\nobject detection and recognition benchmarks. These, along with a better\nunderstanding of deep learning methods, have also led to improved capabilities\nof machine understanding of faces. CNNs are able to detect faces, locate facial\nlandmarks, estimate pose, and recognize faces in unconstrained images and\nvideos. In this paper, we describe the details of a deep learning pipeline for\nunconstrained face identification and verification which achieves\nstate-of-the-art performance on several benchmark datasets. We propose a novel\nface detector, Deep Pyramid Single Shot Face Detector (DPSSD), which is fast\nand capable of detecting faces with large scale variations (especially tiny\nfaces). We give design details of the various modules involved in automatic\nface recognition: face detection, landmark localization and alignment, and face\nidentification/verification. We provide evaluation results of the proposed face\ndetector on challenging unconstrained face detection datasets. Then, we present\nexperimental results for IARPA Janus Benchmarks A, B and C (IJB-A, IJB-B,\nIJB-C), and the Janus Challenge Set 5 (CS5).\n