Humans and algorithms for facial recognition: the effects of candidate list length and experience on performance

Abstract

These experiments investigated how the candidate list displayed to facial reviewers affects their performance in a one-to-many unfamiliar face matching task. Automated facial recognition systems present the results of a database search and require selection of an image that matches the target. Few studies investigate how humans in combination with facial recognition algorithms perform within different operational contexts. These experiments investigated how the candidate list displayed to facial reviewers affects their performance in a one-to-many unfamiliar face matching task. We tested candidate list length with inexperienced (Experiment 1) and experienced (Experiment 2) facial reviewers. Candidate list length had a large impact on performance, varying with the operational context. However, response-time analyses show that the accurate responses were resolved quickly, with an error-prone guess process implemented after failed search. Long candidate lists (100 images) produced more false alarms, fewer hits, lower decision confidence, and increased response latencies among both inexperienced and experienced facial reviewers.