False alarm rate: a critical performance measure for face recognition

The performance of a face recognition algorithm is typically characterised by correct identification rate under the closed-world assumption. To be of greatest practical use, the closed-world assumption must be relaxed and the classifier used both for detection and identification. It is put forward that for open-world applications, the false alarm rate of the classifier is at least as important as the identification rate. Under a repeated verification model, all face recognisers exhibit a rapid non-linear increase in false alarm rate with the false alarm rate of the one-to-one verification used. If the one-to-one false alarm rate is not strictly controlled, the overall classifier are unusable. A method is presented to predict the false alarm rate of a large gallery classifier using only a small data set. It is then shown that the false alarm error rate is always greater than the identification error rate. Therefore the false alarm rate is a more difficult criterion to minimise when designing a classifier.