Figure-Ground Separation By Contour Statistics

Vision is the very process that reinterprets images as physical scenes, such as separatingthe figure out of an image. As all other vision related tasks, figure-ground separationdoes not receive enough information from the image itself, but requires priors to helpsettling down the final segmentation.The Gestalt Laws for Perceptual Organization is the very prior that facilitatesfigure-ground separation by constraining biological statistics of the foreground contour.However, quantitative research about the law is limited because of the lack of acomputational model that is biological plausible.In this paper we propose a Markov Random Fields based representation for the GestaltLaw, and suggest using a Message Passing-like scheme to infer the segmentation. TheMRF function is specially encoded to consider orientations along the contour, thus theGestalt law is embedded into the inference. In contrast, previous segmentation modelssuch as Grabcut seldom consider such priors if ever.As a basic framework, our system is designed in reference to neurophysiological results.Its architecture is composed of three loosely-coupled modules. The first module, withregard to primal visual cortex(V1), is used to extract simple geometric features such asedges and corners. Based on these features, the second one, say the 'extra-striate cortexV2', infers the perceptual contour. The third module is used to select the interestedregion like higher cortical areas do. In our paper, this module is replaced by apre-attentive model or a interactive interface. This architecture is consistent to brainscience and has a good perspective for expansion.We also studied the forward interaction between the V1 and V2 area. In this step, wequantitatively modeled the Gestalt Laws. Difference from other quantitative works, wefocused on a neuronal implementation of the Law. More specifically, our design evencovers 'the law of closure', which was not covered by any of previous works, but wasessential for figure-ground perception.To validate our method, we conducted experiments on two psychological figures andthe Berkeley Segmentation Dataset. The results are inspiring but reflect limitation foronly including the Gestalt Law. We also proposed an interactive segmentationalgorithm to provide more useful solutions.