Ground truth and evaluation for latent fingerprint matching

Abstract

In forensic fingerprint studies annotated databases is important for evaluating the performance of matchers as well as for educating fingerprint experts. We have established ground truths of minutia level correspondences for the publicly available NIST SD27 data set, whose minutia have been extracted by forensic fingerprint experts. We performed verification tests with two publicly available minutia matchers, Bozorth3 and k-plet, yielding Equal Error Rates of 36% and 40% respectively, suggesting that they have similar (poor) ability to separate a client from an impostor in latent versus tenprint queries. However, in an identification scenario, we found performance advantage of k-plet over Bozorth3, suggesting that the former can rank the similarities of fingerprints better. Regardless of the matcher, the general poor performance is a confirmation of previous findings related to latent vs tenprint matching. A finding influencing future practice is that the minutia level matching errors in terms of FA and FR may not be balanced (not equally good) even if FA and FR have been chosen to be so at finger level.

Pre-registration for Improved Latent Fingerprint Identification

“…If the distance between a minutia from ideal and matched sets is less than 12 pixel units, then they are assumed to be corresponding mated pairs. A detailed study on NIST-SD27 where these kind of discrepancies between the ideal and matched minutia sets is reported in [14], where a refined version of the ground truth minutia sets for NIST-SD27 is made publicly available [14].…”

SAFE features for matching fingermarks by neighbourhoods of single minutiae

“…Further experiments on multiple minutiae are required for benchmarking present work. The observed EER using a publicly available (source code) minutiae-only matcher shows 36% EER [16] performance on an ideal set (and 6% on a matched set -which is an unrealistic scenario), which is significantly poorer than the achievement of the suggested method.…”

“…We have selected 1-2 points from every image in areas of significant orientation change (in neighbourhood of core or deltas if present) resulting in 320 clients and 50,978 impostors. Because we are testing the descriptive power of features at a single point, without using the constellation of minutiae, such minutia are good candidates for testing descriptive power [16]. Annotations of the expert for location and direction of the minutiae were used when testing our features.…”

Asynchronous Processing for Latent Fingerprint Identification on Heterogeneous CPU-GPU Systems

“…Many studies analyze the performance of general identification algorithms in processing latent fingerprint databases. The achieved results revealed a poor performance owing to the low quality of the input data [30] and thus, opening the way to the development of new algorithms specifically designed to this aim. In latent fingerprint identification, early works proposed several solutions for handling typical deformations which affect the matching procedure.…”