The major contribution of this paper is a duality between the identification and verification evaluation protocols. The duality (1) gives a mapping between identification and verification scores and (2) bounds identification and verification scores in terms of each other. In the identification--verification duality, the roles of the gallery and probes set are reversed between the two problems, and a mapping between the axis of the cumulative match curve and ROC is established. Our identificaton--verification duality thereom states that ROCs are lower bounds for cumulative match curves. We also show that the bound is tight. We use the duality bound to characterize the effect of increasing gallery size on identification performance. Our results are empirically checked using the Sep96 FERET test results for a baseline principal components analysis (PCA)-based face recognition algorithm.
Proceedings of Auto ID
Biometrics, identification, ROC, verification
On Performance Statistics for Biometric Systems, Proceedings of Auto ID
(Accessed December 1, 2023)