I. Introduction
Some stereo image-matching methods require a user-selected set of corresponding points (seed points) in the left and right images to initiate automated stereo matching routines [1]. For example, Tang et al. [6] presented a dense matching method by making use of a Voronoi diagram constructed by reliably matched seed points, with the whole image being divided into cells, each cell containing a seed point taken for propagation inside its region, and the eight corresponding neighboring points of each seed are found using the disparity of this centered point under the continuity constraint. Chen et al. [2] utilized pyramidal stereo matching for urban three-dimensional building modeling; they divided the original image into several areas, and selected the most reliably matched points in each area as the seed points of the up-layer image matching. Zhu et al. [8] presented a robust image-matching propagation method based on the self-adaptive triangle constraint, which relies on the initial corresponding triangulations formed by a few seed points in the stereo pairs.