Comparison of projection-based face recognition algorithms

Traditional methods of evaluating face recognition algorithms report on their performance on a single set of images. However, this does not fully capture algorithm performance or allow for statistical comparisons between algorithms. To overcome this problem, we present an evaluation methodology that allows one to obtain a robust and comprehensive assessment of face recognition algorithms. From the evaluation methodology, we formulate a model that allows for statistical hypothesis testing. We use this to compare the performance of different algorithms of the same test set, and one algorithm on different test sets.