Computers do better than experts matching faces in a large population

Abstract

The performance of computer-based face recognition algorithms improves significantly recently. Although extensive evaluations of algorithms have been carried out, there has been little work on measuring the performance of experienced humans matching faces. We compared the face verification performance of 4504 experienced inspectors with a leading face recognition algorithm in a large population. Experts and algorithms determined whether face image pairs taken under real-world situation were pictures of the same person or not. The face recognition algorithm evidently surpassed experts on “easy” and “middle” face pairs, while on “hard” ones experts were superior to the algorithm. As a whole the algorithm outperformed experts on the face matching task. The distinct performance of the algorithm and experts on unfamiliar faces underscores the need to understand how much humans benefit from experience. It also suggested utilizing human-machine partnerships in security applications. A practical cascade system was proposed to reduce the workload of inspectors and achieve accuracy better than both experts and algorithms.