Unconstrained face detection: State of the art baseline and challenges

Abstract

A large scale study of the accuracy and efficiency of face detection algorithms on unconstrained face imagery is presented. Nine different face detection algorithms are studied, which are acquired through either government rights, open source, or commercial licensing. The primary data set utilized for analysis is the IAPRA Janus Benchmark A (IJB-A), a recently released unconstrained face detection and recognition dataset which, at the time of this study, contained 67,183 manually localized faces in 5,712 images and 20,408 video frames. The goal of the study is to determine the state of the art in face detection with respect to unconstrained imagery which is motivated by the saturation of recognition accuracies on seminal unconstrained face recognition datasets which are filtered to only contain faces detectable by a commodity face detection algorithm. The most notable finding from this study is that top performing detectors still fail to detect the vast majority of faces with extreme pose, partial occlusion, and/or poor illumination. In total, over 20% of faces fail to be detected by all nine detectors studied. The speed of the detectors was generally correlated with accuracy: faster detectors were less accurate than their slower counterparts. Finally, key considerations and guidance is provided for performing face detection evaluations. All software using these methods to conduct the evaluations and plot the accuracies are made available in the open source.