Abstract
Human face recognition is more accurate for faces of one's own race than for faces of other races. The purpose of this study was to determine whether face recognition algorithms show an “other-race effect”. We tested 13 algorithms from a recent international competition: eight from Western countries (France, Germany and the United States) and five from East Asian countries (China, Korea, and Japan). The algorithms were required to match facial identity in pairs of images (a controlled illumination image and an uncontrolled illumination image). We first assessed algorithm performance on Caucasian (n = 3,359,404) and East Asian (n = 205,114) face pairs at the low false alarm rates required for security applications. Algorithm performance was measured by fusing the East Asian algorithms and the Western algorithms separately. The Western fusion algorithm recognized Caucasian faces more accurately than East Asian faces and the East Asian fusion algorithm recognized East Asian faces more accurately than Caucasian faces. Next, we carried out a direct comparison between humans of Caucasian and East Asian descent and the face recognition algorithms. In this case, we used a manageable number of face pairs (40 East Asian and 40 Caucasian pairs) and employed a more general test that considered performance across the full range of false alarms. For humans, we found the standard other-race effect. However, both the East Asian and Western fusion algorithms performed better on Caucasian faces—the “majority” race in the database used in the competition. The performance advantage for Caucasian faces was substantially larger for the Western fusion algorithm than for the East Asian fusion algorithm. We discuss these results in the context of the short-term and long-term perceptual tuning (algorithm training) that may underlie the pattern of results. We conclude that algorithms can show the other-race effect under some conditions.
Work funded by the Technical Support Working Group (TSWG).