IARPA Janus Benchmark Multi-Domain Face

Face recognition performance degrades when images are captured in difficult environments and non-ideal imaging conditions, for example in surveillance video. Heterogenous face recognition applications that require matching of images across varying distances, resolution, and visible and infrared imaging bands pose additional challenges. Deep-learning techniques have been employed to improve unconstrained face recognition performance, but they require a very large amount of data. There is a distinct need for publicly-available facial video datasets, collected with the consent of the subjects, that can be used to overcome the challenge of performing facial recognition in multi-domain, heterogenous scenarios. This paper presents a multi-domain face image dataset comprised of ground truth and operational images and video captured using a variety of cameras. These include fixed and body-worn cameras capable of imaging at visible, short-mid-, and long-wave infrared wavelengths, including images captured at distances up to 500m. Over 161 hours of video were collected from 251 subjects. This paper provides a description of the data collection, post processing, and baseline matching performance of this unique dataset.