Abstract
Figure-ground separation can be driven by differences in luminance, contrast, orientation, and other local cues. In natural scenes, these multiple cues are intertwined. To probe how they interact, we constructed a space of synthetic textures in which these cues could be separately manipulated (Vision Res. 2015). We used these textures to create images in which the cues that defined figure and ground were varied independently. We then asked whether figure-ground separation is driven simply by the difference between figure and ground, or rather, whether the compositions of figure and ground also play a role. We focused on textures defined by their second-order statistics, as these contained both contrast and orientation information. Four orientations (cardinal and oblique) were studied.
Subjects (N=3) carried out a 2-AFC task, identifying a target image that contained five randomly-positioned circular figures (25% of the total area) defined by one set of local image statistics, superimposed on a background defined by a different set of image statistics. The non-target image was statistically uniform, and matched the target image’s statistics averaged across space.
Thresholds for figure-ground separation depended not only on the figure-ground difference, but also on their individual contents. However, the balance of these factors depended on the extent to which the textures were oriented. For textures that were blob-like (e.g., positive correlations on both horizontal and vertical axes), only the figure-ground difference mattered. For textures that were strongly oriented (e.g., positive correlations on the horizontal axis but negative correlations on the vertical axis), the composition of figure and ground had a large effect, influencing threshold by up to a factor of two even when figure-ground differences were held constant.
In sum, figure-ground separation makes use not only of texture differences, but also of the orientation composition of figure and ground.