Stereo vision matching over single-channel color-based segmentation

Stereo vision is one of the most important passive methods to extract depth maps. Among them, there are several approaches with advantages and disadvantages. Computational load is especially important in both the block matching and graphical cues approaches. In a previous work, we proposed a region growing segmentation solution to the matching process. In that work, matching was carried out over statistical descriptors of the image regions, commonly referred to as characteristic vectors, whose number is, by definition, lower than the possible block matching possibilities. This first version was defined for gray scale images. Although efficient, the gray scale algorithm presented some important disadvantages, mostly related to the segmentation process. In this article, we present a pre-processing tool to compute gray scale images that maintains the relevant color information, preserving both the advantages of gray scale segmentation and those of color image processing. The results of this improved algorithm are shown and compared to those obtained by the gray scale segmentation and matching algorithm, demonstrating a significant improvement of the computed depth maps.