Abstract
Vision science has experienced a recent surge in theories that model perceptual organization as a form of Bayesian inference. Following Helmholtz's principle of unconscious inference, an underlying assumption of such approaches is that perceptual organization is treated as the output of processes that derive the most plausible interpretation of the image data given a set of constraints (or priors). An alternative view, articulated most forcefully by the Gestalt psychologists, emphasizes the role of dynamical principles of organization to explain processes of perceptual organization. Here, I present a class of illusory contours that defy explanation in Bayesian frameworks. A class of robust illusory contours are reported which reveal that contours will form even when they play no necessary part in explaining image data. In one set of displays, a moving outline figure is occluded by a visible surface. Although the pattern of occlusion and disocclusion is fully explained by the visible surfaces in the scene, additional vivid illusory contours are formed that appear to move and deform. A series of experiments reveal that the strength of these contours is completely determined by the relative contrasts of the occluding contour terminator, and the occluding surface relative to the background. This contrast dependence demonstrates that such contours emerge from relatively low level processes that arise early in the visual processing stream. I show that these illusory contours form either in the presence of visible occluding surfaces (i.e., those generating suprathreshold values of image contrast), or when the occluding surface is specified by another illusory contour (i.e., two independent illusory contours can form simultaneously, even when either one alone would fully explain the image data). It is argued that these phenomena challenge the explanatory power of Bayesian models of contour synthesis, and perception more generally.