BioHDD: a dataset for studying biometric identification on heavily degraded data

Substantial efforts have been put into bridging the gap between biometrics and visual surveillance, in order to develop automata able to recognise human beings `in the wild´. This study focuses on biometric recognition in extremely degraded data, and its main contributions are three-fold: (1) announce the availability of an annotated dataset that contains high quality mugshots of 101 subjects, and large sets of probes degraded extremely by 10 different noise factors; (2) report the results of a mimicked watchlist identification scheme: an online survey was conducted, where participants were asked to perform positive and negative identification of probes against the enrolled identities. Along with their answers, volunteers had to provide the major reasons that sustained their responses, which enabled the authors to perceive the kind of features that are most frequently associated with successful/failed human identification processes. As main conclusions, the authors observed that humans rely greatly on shape information and holistic features. Otherwise, colour and texture-based features are almost disregarded by humans; (3) finally, the authors give evidence that the positive human identification on such extremely degraded data might be unreliable, whereas negative identification might constitute an interesting alternative for such cases.