Transformation, Ranking, and Clustering for Face Recognition Algorithm Comparison

The performance of face recognition algorithms is recently of increased interest. Empirical analyses of algorithms have traditionally been limited to rank-based scores such as cumulative match and receiver operating characteristics. This restriction to performance measures based on rank-based statistics arises because it is not possible to directly compare similarities output by algorithms. This paper presents the Phi-PIT transformation that makes it possible to compare such heterogeneous outputs, and allows a large body of classical statistical methods to be used to measure and analyze performance. These statistical techniques inclucde multiple comparison techniques, analysis of variance (ANOVA), and regression techniques. This paper presents ANOVA, graphical ANOVA, and StudentNewman-Keuls clustering analyses of the transformed outputs of fifteen face recognition algorithms.