Factors that influence algorithm performance in the Face Recognition Grand Challenge

• Published in: Computer vision and image understanding Vol. 113; no. 6; pp. 750 - 762
• Main Authors: Beveridge, J. Ross, Givens, Geof H., Phillips, P. Jonathon, Draper, Bruce A.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Inc 01.06.2009; Elsevier
• Subjects: Applied sciences; Artificial intelligence; Computer science; control theory; systems; Data processing. List processing. Character string processing; Exact sciences and technology; Face recognition; Memory organisation. Data processing; Pattern recognition. Digital image processing. Computational geometry; Performance analysis; Software; Statistical modeling; Subject covariates; Statistical modeling; Subject covariates; Performance analysis; Face recognition; Performance evaluation; Computer vision; Generalized linear model; Image resolution; Statistical analysis; Probabilistic approach; Image processing; Sex; Pattern recognition; Modeling; Mixed model; Facies; Age
• ISSN: 1077-3142; 1090-235X
• DOI: 10.1016/j.cviu.2008.12.007

A statistical study is presented quantifying the effects of covariates such as gender, age, expression, image resolution and focus on three face recognition algorithms. Specifically, a Generalized Linear Mixed Effect model is used to relate probability of verification to subject and image covariates. The data and algorithms are selected from the Face Recognition Grand Challenge and the results show that the effects of covariates are strong and algorithm specific. The paper presents in detail all of the significant effects including interactions among covariates. One significant conclusion is that covariates matter. The variation in verification rates as a function of covariates is greater than the difference in average performance between the two best algorithms. Another is that few or no universal effects emerge; almost no covariates effect all algorithms in the same way and to the same degree. To highlight one specific effect, there is evidence that verification systems should enroll subjects with smiling rather than neutral expressions for best performance.