Abstrakt
Creating a lineup of a perpetrator of another ethnicity is a challenging task. It is often impossible to compile such live lineups with a sufficient number of fillers where no person stands out.
Therefore, a photo array is likely to be picked over a live lineup, but still may face challenges due to a lack of available resources and time constraints. To simplify the administrative burden on police officers, we created a prototype application which uses an inter-person similarity metric based on both a structured description of persons and visual descriptors received from a deep convolutional neural network.
The outcomes of this program were compared to police lineups in our research on lineup fairness using mock-witnesses (i.e. participants who have not seen the perpetrator, just obtained his description). A total of 864 identifications were made.
Both the police and computer-made lineups had similar results towards a Caucasian ethnicity suspect. The computer had a slightly worse outcome for the Asian suspect, but it completed the task nearly instantaneously.
Therefore, we see a great potential in "human-in-the-loop" solutions, where suggestions from computer methods are being confirmed by police technicians. This work was supported by the Czech grant GAUK-232217.