A model assigning 3D surface-ownership

Abstract

It is known that human visual system has the mechanism of figure-ground separation distinguishing foreground object and background regions. The border-ownership coding, represented by neurons having the selectivity for figural side against contour, in addition to orientation selectivity, is thought as the internal representation of the figure-ground relation in the brain. However, we usually look 3D environment binocularly, and the boundary of objects is 3D surface, rather than 2D contour. It is necessary to extend the nature of the border-ownership from 2D to 3D, where each local surface patch belongs to one side between two sides of adjacent subspaces. How such information processing is realized comes to a problem. This study presents a neural network model assigning the ownership of 3D surface. We focus on the nature that alike 2D object’s boundaries as contours, 3D object’s boundaries as surfaces have globally convex shapes, though objects have generally both convex and concave parts. We extended 2D figure-ground separation model proposed by Kikuchi and Akashi (2001) into 3D model with similar principle. We supposed the signal propagation by local averaging equipped in each two sides against the surface, and mutual inhibition between the two sides. Initial values in proportion to the magnitude of curve is given to the inner side of the curve. We confirmed that through the iterative process signals remains in only one side corresponding to the inner region of object.