Abstract
Different locations in the visual environment vary greatly in terms of how likely they are to draw a person's attention. When inferring the most likely target of another person's gaze, it would therefore be a reasonable strategy to incorporate expectations about the relative visual salience of these various locations, weighing this prior knowledge against incoming social signals (e.g., eye cues). This Bayesian approach to modeling gaze perception has informed computer vision techniques, but whether this model accounts well for human performance remains an untested hypothesis. We present subjects with a "gazer" fixating his eyes on various locations on a two-dimensional surface, and project arbitrary photographic images onto that surface. Subjects judge where the gazer is looking in each image. A full Bayesian model, which takes image salience information into account, fits subjects' gaze judgments better than a reduced model that only considers the perceived direction of the gazer's eyes. Varying the amount of time the subject is allowed to view the gazer reveals that center bias tends to dominate gaze judgments early, whereas salient features specific to the projected image influence judgments at longer viewing durations.