Abstract
Since early gaze-tracking experiments, a wealth of studies have sought to estimate where people look in natural images. Typically, the data for these studies come from free-viewing experiments that share similar settings. A consistent finding in these studies is that eye-fixations fall more often at or near the center of the image than its periphery. Multiple reasons have been examined for contributing to this center-bias, including the image position on the screen and feature biases in the dataset images. However, the stimuli in these studies encapsulate multiple types of centers, all of which coincide at the same spatial location and therefore are indisciminable. The reason for this is that virtually always the stimuli are perceptually framed very simplistically, namely within a rectangle. In order to determine which of these different types of centers is responsible for the center-bias, we manipulate the boundaries of the stimuli into different shapes, including non-convex and/or asymmetric polygons. This allows us to specify different types of centers (and in particular, the center of the convex-hull, the center of the bounding-box, the center of mass, and the visual center) and systematically estimate the tendency towards each of them. Our analysis unveils the influence of stimulus boundaries on the center-bias and suggests that the center of mass of the stimulus is the one that facilitates the content-independent center-bias. This finding also supports the hypothesis that observers tend to fixate more often at the center of mass of visual objects.