Predicting Human Performance for Face Recognition

Abstract

The ability of humans to recognize faces provides an implicit benchmark for gauging the performance of automatic face recognition algorithms. IntroductionHuman face recognition abilities are impressive by comparison to the performance of many current automatic face recognition systems. Such is the belief of many psychologists and computer scientists. This belief, however, is supported more by anecdotal impression than by scientific evidence. In fact, there are relatively few systematic and direct comparisons between the performance of humans and automatic face recognition algorithms, (though see [84], for exceptions). Comparisons of these sorts can be carried out both at a quantitative and qualitative level. On the qualitative side, as the field of automatic face recognition develops, the assessment of working systems has expanded to consider, not only overall levels of performance, but performance under more restricted and targeted conditions. It may thus be important to know, for example, whether an algorithm performs accurately recognizing "male faces of a certain age", "female faces of a certain ethnicity", or faces that are "typical" in the context