Uncertainty-aware Comparison Scores for Face Recognition

Estimating and understanding uncertainty in face recognition systems is receiving increasing attention as face recognition systems spread worldwide and process privacy and security-related data. In this work, we investigate how such uncertainties can be further utilized to increase the accuracy and therefore the trust of automatic face recognition systems. We propose to use the uncertainties of extracted face features to compute a new uncertainty-aware comparison score (UACS). This score takes into account the estimated uncertainty during the calculation of the comparison score, leading to a reduction in verification errors. To achieve this, we model the comparison score and its uncertainty as a probability distribution and measure its distance to a distribution of an ideal genuine comparison. In extended experiments with three face recognition models and on six benchmarks, we investigated the impact of our approach and demonstrated its benefits in enhancing the verification performance and the genuine-imposter comparison scores separability.