Abstract
Algorithmically-defined visual textures provide a way to analyze computations performed in early vision. In this approach, image sets are synthesized that contain independent, controlled variation of several types of local features. The method therefore allows for a characterization of processing of individual feature types and how they interact – something difficult to accomplish with natural images. Here we extend this strategy from elementary visual features to shape. The starting point is our texture-based analysis of contrast, edge, and corner. Here, we used binary textures: each check was black or white according to an algorithm that controlled the one-, two-, three-, and four-point statistics within 2x2 neighborhoods of checks. To extend this idea to elements of shape, we replace the uniformly black or white checks by tiles containing curved segments. The tiles are designed so that where they meet, they form continuous contours. This construction produces an image with extended shapes, whose characteristics are controlled by the underlying binary texture. Specifically, manipulating the two-point and higher-order statistics of this underlying texture produces new textures with curvilinear contours and varying amounts of convex and non-convex shapes. These new textures are balanced for lower-level features such as the length, orientation, and curvature of individual segments. To use these stimuli to analyze texture segmentation based on shape, we asked subjects (N=6) to perform a 4-AFC texture segmentation paradigm, adapted from previous studies. Textures containing isolated circles were more readily segregated than textures containing the equivalent amount of curved segments, but combined in a way that did not produce isolated shapes. Further, simple circles were a stronger segmentation cue than more complex shapes. Because of the construction of the textures from identical sets of tiles, the superiority of circles could not be explained on the basis of the curvature of local segments.
Meeting abstract presented at VSS 2016