Correspondence from color shading

Presents a region-based matching algorithm, working on the color shading information (red, green and blue) contained in an image pair. In each color image, the regions are defined as connected areas within which the shading surface (i.e. the surface whose height at each location equals the color brightness in the image) maintains constant-sign Gaussian and mean curvatures. Region characteristics and adjacency information are explored by means of a region graph, in which each node corresponds to a particular region and is labeled with a vector of intraregional characteristics, while the arc joining two adjacent nodes is labeled with the relative position of the corresponding regions. The search for correspondences is restricted by means of a curvature consistency constraint, which states that matching regions must be embedded in areas of the shading surfaces having similar curvature characteristics. The disparity maps extracted from the R, G and B signals are combined at each resolution level of a coarse-to-fine strategy.<>