What do quality measures predict in biometrics?

This paper is discusses the role of quality measures in biometric classification. We challenge a common notion that quality measures are performance predictors of the baseline biometric classifier. Instead, we postulate that quality measures are class-independent classification features, and as such are conditionally relevant class predictors. We present a systematic, probabilistic approach towards error prediction in biometric classification systems, where quality measures play an integral role in a stacked classifier ensemble. We demonstrate the results of error prediction in face verification using the proposed method.