Generalized Linear Mixed Model Evaluation of Face Verification Algorithms, Emphasizing Human Subject Characteristics

• Activity Number: 20
• Type: Topic Contributed
• Date/Time: Sunday, August 7, 2005 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statisticians in Defense and National Security
• Author(s): Geof Givens*+ and Ross Beveridge and Bruce Draper
• Companies: Colorado State University and Colorado State University and Colorado State University
• Keywords: biometrics ; face recognition ; verification ; GLMM

Abstract

We begin with a brief review of the human face verification problem and related statistical algorithms. One of the most vexing problems facing algorithm designers is that some faces are simply harder to verify than others and it is difficult to isolate the reasons. We present an empirical analysis of how verification success relates to subject age, gender, facial hair, and the status of subjects' eyes and bangs. Using a generic PCA verifier, 1072 pairs of subject images from the FERET database were run though verification experiments at seven false positive tolerances. The response was log odds of correct verification. The fitted model allowed us to quantify the relative magnitudes of components of variation in verification difficulty. Significant effects for many of the covariates were confirmed. Furthermore, the effect of the false positive rate tolerance was loglinear and exhibited significant interactions with the subject-based predictors. These results have implications for further research, the ongoing Face Recognition Grand Challenge, and algorithm development and evaluation.