This paper introduces linear models (LM), generalized linear models (GLM), and generalized linear mixed models (GLMM) for analyzing performance of face recognition algorithms. These three statistical techniques are applied to analyzing the affect of subject attributes on performance. To support this analysis, subject covariate data were collected on 1,072 pairs of FERET images corresponding to 1,072 human subjects in order to study how these covariates influence the identification and verification performance of standard face recognition algorithms. Up to 11 factors are included in these studies: race (white, asian, black, or other), gender, age (young or old), glasses (present or absent), facial hair (present or absent), bangs (present or absent), mouth closed or other), eyes (open or other), complexion (clear or other), makeup (present or absent), and expression (neutral or other). Linear models are used to relate similarity scores to subject covariates for a standard Principle Components Analysis (PCA)algorithm, an Interpersonal Image Difference Classifier (IIDC) and an Elastic Bunch Graph Matching (EBGM) algorithm. A generalized linear model is used to predict probability of rank one recognition for the same set of algorithms. Finally, a generalized linear mixed model is used to predict the probability of correct verification for the PCA algorithm at different false acceptance rates. Two questions underly these studies. First, can LMs, GLMs, and GLMMs be successfully to analyzing face recognition performance? Second, what factors or combinations of subject factors make automated face recognition difficult? The answer to the first question is yes. The second question is answered in this paper.
Conference Dates: April 10-12, 2006
Conference Location: southhampton,
Conference Title: IEEE 7th International Conference on Automatic Face and Gesture Recognition
Pub Type: ConferencesReport Number:
biometrics, evaluation, face recognition, generalized linear models, linear models