Intrinsic Limitations of Fingerprint Orientation Estimation

Abstract

Estimation of orientation field is a crucial issue when processing fingerprint samples. Many subsequent fingerprint processing steps depend on reliable and accurate estimations. Algorithms for such estimations are usually evaluated against ground truth data. As true ground truth is usually not available, human experts need to mark-up ground truth manually. However, the accuracy and the reliability of such mark-ups for orientation fields have not been investigated yet. Mark-ups produced by six humans allowed insights into both aspects. A Root Mean Squared Error of about 7 against true ground truth can be achieved. Reproducibility between two mark-ups of a single dactyloscopic expert is at the same precision. We concluded that the accuracy of human experts is competitive to the best algorithms evaluated at FVC-ongoing.