Binocular model for figure-ground segmentation in translucent and occluding images

A Fourier-based solution to the problem of figure-ground segmentation in short baseline binocular image pairs is presented. Each image is modeled as an additive composite of two component images that exhibit a spatial shift due to the binocular parallax. The segmentation is accomplished by decoupling each Fourier component in one of the resultant additive images into its two constituent phasors, allocating each to its appropriate object-specific spectrum, and then reconstructing the foreground and background using the inverse Fourier transform. It is shown that the foreground and background shifts can be computed from the differences of the magnitudes and phases of the Fourier transform of the binocular image pair. While the model is based on translucent objects, it also works with occluding objects. (C) 2002 society of Photo-Optical instrumentation Engineers.