Topic: Utility-centric Ranking of Quality Assessment Methods
by Torsten Schlett
online Big Blue Button Room: D19/2.03a, June 23, 2022 (Thursday), 12.00 noon
Keywords — Evaluation Methodology, Sample Quality, Biometric Utility, Error vs Discard Characteristic, Face Recognition
The quality of biometric samples can vary in terms of the samples‘ utility for biometric comparisons, and automatic quality assessment methods exist. How can we best quantify how well these quality assessment methods perform relative to each other? This talk investigates the stability of method rankings produced by pAUC (partial Area Under Curve) values for the common EDC (Error vs Discard Characteristic) evaluation approach, as well as a simple correlation coefficient ranking approach. Concrete examples and experiments with real data are shown for FIQA (Face Image Quality Assessment) specifically, but the talk can apply to the ranking of quality assessment methods in general, and experiments with fully synthetic scores are also included.