Dense Stereo Correspondence Using Polychromatic Block Matching

Only few problems in computer vision have been investigated more vigorously than stereo. Nevertheless, almost all methods use only gray values and most of them are feature-based techniques, i.e., they produce only sparse depth maps. This paper presents an efficient technique for dense stereo correspondence using a new Polychromatic Block Matching. Four different color models (RGB, XYZ, IlI2I 3, HSI) and three different color measures have been investigated with regard to their suitability for stereo matching. As a result the IlI2I 3 color space provides the best information for stereo when using the Euclidean distance for color measurement. 1 Introduct ion and Mot ivat ion Worldwide many research activities dealing with stereo vision are known. Nevertheless, almost all authors use only gray value images. The results are rather acceptable, but higher precision is required for the reconslxuction of visible surfaces. There are two ways to improve the results. One approach is to seek new mathematical techniques; another approach -as is done here -is to seek a more complete and efficient use of available image information. That is the analysis of color in stereo images. There are several motivations for using chromatic information. First, chromatic information is easily obtained with high precision when using a 3-chip CCD camera. Second, color plays an important role in human perception. Third, it is obvious that red pixels cannot match with blue pixels although their intensities are equal or similar. Last not least, the existing computational approaches to color stereo correspondence have shown that the matching results can be considerably improved when using color information. Most color stereo techniques are feature-based [1, 2, 3], i.e., only scattered control points can be computed for the succeeding surface reconstruction process. Therefore, there is an essential need for algorithms that compute dense disparity maps defined for every pixel in the entire image. Unfortunately, the dense color stereo techniques are very time consuming because they use simulated annealing [4] or a statistical criterion and an iterative technique with a priori unknown convergence speed [5]. In this paper a very efficient algorithm for dense color stereo matching is introduced. The Block Matching technique is chosen for extension to color because of its efficiency already shown for gray value images [6]. So far, all methods mentioned above only use the RGB color space. In addition, it will be shown that the precision of color stereo matching is improved when a suitable color coordinates system and/or color measure is chosen.