Estimating individuality in feature point based retina templates

The lack of large public retina image databases means it is difficult to judge the relative merits of the retina biometric itself, of different scoring functions, or of potential biocryptographic constructs. We derive conservative theoretical genuine and imposter score distributions for feature point based retina templates by normal kernel density estimation. We base them on 147 images from the VARIA database and use 7 scoring functions. This allows us to infer EERs in the range 0.3%-1.3% and, for FNMRs of less than 10%, entropy estimates between 65 bits and 200 bits.