Ecological Statistics of Contour Grouping

The Gestalt laws of perceptual organization were originally conceived as qualitative principles, intrinsic to the brain. In this paper, we develop quantitative models for these laws based upon the statistics of natural images. In particular, we study the laws of proximity, good continuation and similarity as they relate to the perceptual organization of contours. We measure the statistical power of each, and show how their approximate independence leads to a Bayesian factorial model for contour inference. We show how these local cues can be combined with global cues such as closure, simplicity and completeness, and with prior object knowledge, for the inference of global contours from natural images. Our model is generative, allowing contours to be synthesized for visualization and psychophysics.