Abstract
The Münsterberg (Café wall) illusion belongs to a class of illusions of tilt whose effectiveness crucially depends on the luminance polarities of the elements of the displays. A computational model of such phenomena was constructed, in the form of a network of units which simulate simple cells in V1. The receptive fields of the units were modeled as Gabor functions or as differences of shifted Gaussians, with different sizes and orientations. Inputs for the model were matrices of 32x32 units arranged on a 2D grid, representing various visual stimuli, and outputs were 32x32 matrices of reactions of network units to the stimuli. The model was tested with two classes of 'control' stimuli and two classes of 'experimental' stimuli. One class of control stimuli consisted of simple displays which did not contain tilted elements (such as normally oriented checker-board type patterns), and which did not evoke percepts of tilt from horizontal or vertical; the other class consisted of related displays which did contain tilted elements (tilted or sheared checker-board patterns) and did evoke percepts of tilt. The experimental stimuli were various standard and novel variants of the Café wall configuration and configurations devised by Akiyoshi Kitaoka (such as the 'enhanced checkered illusion'), one class of which evoked strong percepts of tilt, and the other class which generally did not evoke such percepts. The main finding was that the displays that evoked impressions of tilt, whether real (control stimuli) or illusory (experimental stimuli), exhibited similar signature patterns along edges in the simulation outputs, involving characteristic shifted local activity profiles. In contrast, displays that did not evoke tilt impressions lacked such patterns. The conclusion is that the reason that illusory tilt is evoked by this class of displays is that they cause neural activity distributions similar to those caused by actually tilted displays.
Meeting abstract presented at VSS 2014