The Challenge of Face Recognition From Digital Point-and-Shoot Cameras
P J. Phillips, J. R. Beveridge, David Bolme, Bruce A. Draper, Geof H. Givens, Yui M. Lui, Hao Zhang, W T. Scruggs, Kevin W. Bowyer, Patrick J. Flynn, Su L. Cheng
Face recognition is appearing in personal and commercial products at an astonishing rate, yet reliable face recognition remains challenging. Users expect a lot; they want to snap pictures and have their friends, family and acquaintances recognized. This scenario is playing out millions of times a day, and despite its simplicity face recognition in this context is hard. Roughly speaking, failure rates in the 4 to 8 out of 10 range are common. In contrast, using well controlled imagery error rates drop as low as 1 in 1,000. To spur advancement on point-and-shoot face recognition this paper presents a new challenge problem consisting of 9,376 images of 293 people balanced with respect to distance to the camera, alternative sensors, frontal versus not-frontal views, and varying location. Results are presented for two public baseline algorithms and a high-quality commercial algorithm. At false accept rates of 0.001 and 0.01 the best verification rates presented are 0.21 and 0.41 respectively. Pilot studies illustrate covariates analysis, non-match score sensitivity to covariates and an image quality analysis suggesting commonly suggested quality measures dont capture what is making the problem challenging.
IEEE Conference on Computer Vision and Pattern Recognition