Confidence assessment of face recognition results

Abstract

The paper deals with face recognition evaluation methodology. A complex performance comparison of face recognition algorithms has not been established for specific test databases. A basic fact discussed in the paper is limited statistical accuracy of the results. Face recognition results and reported performance increments need not be convincing due to test database properties. The confidence of results should be taken into account in order to reach convincing and reliable test conclusions. In order to express the confidence of the reached results, ROC curves enhanced by confidence intervals are proposed. Enhanced ROC curves are generally applicable without methodological requirements on a test databases.