Labeling edges with a linear network: An integration of low-level vision modules

An important problem in machine vision today is the integration of low level vision modules like color, stereo, motion, and texture. The authors try to integrate these low-level modules along discontinuities, by taking local discontinuities of modules as input and, while using additional local information from the intensity in the neighborhood of the pixel, assign meaningful labels to the edge at the point. They assume the modules' discontinuities are aligned, preferably after being integrated with the intensity edges using Markov random field, so that both the existence and no-existence of a module discontinuity at a point is a meaningful information.