On combining evidence for reliability estimation in face verification

Face verification is a difficult classification problem due to the fact that appearance of a face can be altered by many extraneous factors, including head pose, illumination conditions, etc. A face verification system is likely to produce erroneous, unreliable decisions if there is a mismatch between the image acquisition conditions during the system training and the testing phases. We propose to detect and discard unreliable decisions based on the evidence originating from the classifier scores- and signal domains. We present a method of combining the reliability evidence, nested in a probabilistic framework that allows high level of flexibility in adding new evidence. Finally, we demonstrate on a standard evaluation database (Banca) how the proposed methodology helps in discarding unreliable decisions in a face verification system.