An Other-Race Effect for Face Recognition Algorithms

Published: May 13, 2010

Author(s)

P J. Phillips, Alice J. O'Toole, Abhijit Narvekar, Fang Jiang, Julianne Ayadd

Abstract

Psychology research has shown that human face recognition is more accurate for faces of one�s own race than for faces of other races. In recent years, interest in accurate computer-based face recognition systems has spurred the development of these systems worldwide. We present evidence for an other-race effect in the face recognition algorithms tested in the most recent international competition. We report results for a �Western algorithm� made by averaging eight algorithms from Western countries (France, Germany and the United States) and an �East Asian algorithm� made by average six algorithms from East Asian countries (China, Korea, and Japan). Under test conditions that match the requirements of security applications, the Western algorithm recognized Caucasian faces more accurately than East Asian faces; and the East Asia algorithm recognized East Asian faces more accurately than Caucasian faces. With a more general test of performance, both the East Asian and Western algorithms performed better on the Caucasian faces�the �majority� race in the database of training and test faces used in the international competition. These findings indicate that state-of-the-art face recognition algorithms are not immune to the performance challenges humans struggle with in recognizing �other-race� faces.
Citation: NIST Interagency/Internal Report (NISTIR) - 7666
Report Number:
7666
Pub Type: NIST Pubs

Download Paper

Keywords

face recognition performance, biometrics, face recognition, face recognition algorithm, human face recognition, performance of face recognition algorithms
Created May 13, 2010, Updated February 19, 2017