Learning small gallery size for prediction of recognition performance on large populations

• Prediction model combines hypergeometric distribution model with a binomial model.
• Optimal size of a small gallery is found by an iterative learning process.
• Chernoff and Chebychev inequalities are used to determine small gallery size in theory.
• Experimental results are shown on a challenging data set of fingerprints (NIST-4).

This paper addresses the estimation of a small gallery size that can generate the optimal error estimate and its confidence on a large population (relative to the size of the gallery) which is one of the fundamental problems encountered in performance prediction for object recognition. It uses a generalized two-dimensional prediction model that combines a hypergeometric probability distribution model with a binomial model and also considers the data distortion problem in large populations. Learning is incorporated in the prediction process in order to find the optimal small gallery size and to improve the prediction. The Chernoff and Chebychev inequalities are used as a guide to obtain the small gallery size. During the prediction, the expectation–maximization (EM) algorithm is used to learn the match score and the non-match score distributions that are represented as a mixture of Gaussians. The optimal size of the small gallery is learned by comparing it with the sizes obtained by the statistical approaches and at the same time the upper and lower bounds for the prediction on large populations are obtained. Results for the prediction are presented for the NIST-4 fingerprint database.