Reliability Estimation for Multimodal Error Prediction and Fusion

This paper focuses on the estimation of reliability of unimodal and multimodal verification decisions produced by biometric systems. Reliability estimates have been demonstrated to be an elegant tool for incorporating quality measures into the process of estimating the probability of correctness of the decisions. In this paper we compare decisionand score-level schemes of multimodal fusion using reliability estimates obtained using two alternative methods. Further, we propose a method of estimating the reliability of multimodal decisions based on the unimodal reliability estimates. Using a standard benchmarking multimodal database we demonstrate that the score-level reliability-based fusion outperforms alternative approaches, and that the proposed estimates of multimodal decision reliability allow for an accurate prediction of errors committed by the fusion module.