Abstract
Separating figure from ground is a crucial step in early visual processing. In complex textured images, local analysis of image statistics can provide several kinds of cues: the statistics within the figure, the statistics within the ground, and the differences between them. Here, we attempt to separate these roles. To address the high dimensionality and inter-relatedness of image statistics, we work in a domain of binary (black-and-white) images, and statistics that describe configurations within 2×2 neighborhoods. This domain has 10 dimensions, encompassing luminance, contrast, orientation, and corners; previous work (Vision Res. 2015 and eLife 2014) showed that perceptual distances in this domain corresponded to the informative aspects of natural images. Moreover, the perceptual distances in this space are simple: sensitivities to positive and negative changes in image statistics are nearly identical and isodiscrimination contours are elliptical, and approximately homogeneous throughout the space. We ask how figure-ground separation depends on perceptual distances in this space. Stimuli consisted of five randomly-positioned circular regions containing a “figure” texture, superimposed on a distinct “ground” texture. Figures constituted 25% of the total area. In a 2-AFC task, subjects were asked to distinguish this target stimulus from a non-target stimulus, consisting of a uniform texture whose statistics matched the area-pooled statistics of figure and ground in the target. Trials were grouped in blocks, in which the subject knew the composition of figure and ground. We found (N=3) that perceptual distance between figure and ground textures was an important factor in determining threshold, but not the only factor. Specifically, when the ground texture was negatively correlated, thresholds were 40 to 50% higher than when the ground texture was positively correlated, even when the figure-ground difference was held constant. Thus, ground composition, as well as figure-ground texture contrast, influences thresholds for separation of figure and ground.
Acknowledgement: NIH EY07977