Abstract
In "negative detection" RSVP tasks, subjects detect whether an object is not a member of a prespecified category, such as "modes of transportation." Although detection accuracy is clearly lower than with basic level detection (e.g., "chair"), it is, nonetheless, well above chance. Could such negative detection be achieved for faces, when the task was to detect an unspecified celebrity's face in a stream of non celebrity faces? Subjects also performed negative detection for the categories of animals, clothes, tools, modes of transportation, and plants. Object and face images, all in color, were taken from a Google Image search. All images were depicted on a homogeneous grey background. The non-celebrity face images were taken from websites featuring colored headshots of aspiring actors and business executives. The celebrities were the top 50 most familiar celebrities, as rated by USC undergraduates, from the USC Face and Voice Celebrity Recognition Tests. In the experimenters' judgment, all the images were of high professional quality with no noticeable difference in quality between the images of celebrities and non-celebrities. Subjects viewed RSVP sequences of 32 images. Targets were present in 50% of the sequences but never in the first six or last six positions. For faces, the images were presented at rates of 114, 132, or 150 msec/image. All the objects were shown at a rate of 76 msec/image. On the basis of diagnostic tests for prosopagnosia, each subject was classified as a developmental prosopagnosic (DP) or a Control. Overall error rates on celebrity face detection was well above chance for Controls and DPs (Supplement Fig. 1), with Contols having a lower error rate. Negative detection of objects was much more accurate and virtually identical for Controls and DPs. There is a face familiarity signal that can be detected at RSVP rates.
Meeting abstract presented at VSS 2017