Method to evaluate pose variability in automatic face recognition performance

A key concern in Automatic Face Recognition (AFR) is the decrease of recognition performance as the quality of images decreases. This paper introduces a method to evaluate the impact of face pose variability on face recognition accuracy. Experiments were conducted using three leading commercial face recognition algorithms on data with poses from 0 to ±20 deg in each of the roll, pitch, and yaw directions per subject. Results indicate that roll variations has small effect on performance, while pitch and yaw variations produce a significant increase in error rates. More recent algorithms show better results at low pose variability.