From the laboratory to the real world : Evaluating the impact of impostors, expertise and individual differences on human face matching performance.

Abstract

Evaluating Human Operator face matching performance in applied settings, such as airports, surveillance and access control settings would not only be logistically difficult, but it may not be possible due to many unknowns, such as the presence of impostors. Consequently, Human Operator performance has most commonly been evaluated experimentally, in well controlled laboratory settings. However, the question is, do the results obtained in the well controlled laboratory settings sufficiently reflect, and can they explain what happens in the real world? This applied problem has motivated the principal aim of this research to evaluate the feasibility of extrapolating one-to-one face matching performance findings from laboratory to the real world access control setting, and, in the process, support the development of an ecologically motivated performance evaluation methodology that could be used for future performance assessments, beyond the research reported this thesis. The approach taken to address this aim stemmed from the focus on identity verification or one-to-one face matching task, predominantly performed within access control settings. This focus helped identify numerous factors that may affect face matching performance within access control settings. As a result, this research evaluated the impact of impostor type and frequency, Human Operator expertise and individual differences on one-to-one face matching performance. A preliminary evaluation (Experiment 1) provided important methodological input into subsequent experiments. To address the principal aim, Human Operator face matching performance was first assessed within a simulated live access control setting (Experiment 2) which was subsequently replicated within a laboratory setting (Experiment 3). Experiment 3 also assessed the performance of an automated FR system performance to evaluate the usability of the current methodology beyond only assessing Human Operator performance. From a methodological perspective, this research emphasised the complexities associated with evaluating and understating applied face matching performance. Applied performance may be contingent on interplay of different factors, depending on the considered applied setting. Therefore, it may not be possible to assess and state one single “level” of Human Operator performance that would be relevant to all applied settings and tasks. Instead, Human Operator performance can be assessed in light of the different environmental and task constraints, with the focus on a set of factors. Applied claims need to be appropriately qualified by explaining the exact nature of the face matching task as well as any other factors that may have affected performance. Finally, having considered the impact of frequency and type of impostors, Human Operator expertise and individual differences, the main finding of this research showed that while overall face matching performance in the live and laboratory settings was equivalent, in the live access control setting, Human Operators were more inclined to indicate that two presented stimuli were a match, suggesting a confirmation bias. These findings are discussed in light of previous work.