Elaine M. Newton

Meta-analysis of face recognition algorithms

To obtain a quantitative assessment of the state of automatic face recognition, we performed a meta-analysis of performance results of face recognition algorithms in the literature. The analysis was conducted on 24 papers that report identification performance on frontal facial images and used either the FERET or ORL database in their experiments. The analysis shows that control scores are predictive of performance of novel algorithms at statistically significant levels. The analysis identified three methodological areas for improvement in automatic face recognition. First, the majority of papers report experimental results for face recognition problems that are already solved. Second, authors do not adequately document their experiments. Third, performance results for novel or experimental algorithms need to be accompanied by control algorithm performance scores.

Dependence characteristics of face recognition algorithms

Nonparametric statistics for quantifying dependence between the output rankings of face recognition algorithms are described Analysis of the archived results of a large face recognition study shows that even the better algorithms exhibit significantly different behaviors. It is found that there is significant dependence in the rankings given by two algorithms to similar and dissimilar faces but that other samples are ranked independently. A class of functions known as copulas is used; it is shown that the correlations arise from a mixture of two copulas.