Feature approaches to suspect identification: effect of multiple raters on system performance

Abstract

In suspect identification, facial features of offenders are used to aid witnesses in identifying suspects. Feature approaches have been proposed as an alternative to the mug-shot album for identifying suspects. All feature approaches are based, at least in part, on subjective ratings by police or suspects on such facial features as length of hair: short 1 2 3 4 5 long. There is a strong, widely-held belief among researchers that since people vary considerably in their ratings of any given person's facial features, ratings from at least 10-12 police raters per suspect must be averaged to obtain an accurate estimate that can be used for comparison with witness feature descriptions. It is argued here that this is a myth, untested and unchallenged. Results from five experiments, and theoretical considerations support the contention that one, or at most two, police raters/suspect guarantee good feature system performance. For single-rater systems witnesses examine, depending on the experiment, an average of 15-55 suspect photographs before the actual offenders photograph is displayed (for police files containing 1000 mug-shots). For two-rater systems, witnesses examine an average of 2-12 photographs before retrieving the offender's photograph. More raters do not improve system performance further.

Keywords:

• Suspect identification
• Mug-shots
• Image retrieval
• Computer image retrieval
• Feature retrieval systems