No abstract

No abstract

No abstract

Stereo images are unavoidably distorted by radiometric distortion, even though the stereo images are taken in the controlled conditions. However, the majority of the existing stereo matching methods assume that corresponding pixels in a stereo pair have same intensity values, so their performances degrade significantly using stereo images with radiometric variations. In this paper, we propose a novel local stereo matching method that can operate robustly with stereo images captured with different illuminations and exposures between two cameras. We exploit Kemeny and Snell's distance to compute matching cost values and use segmentation-based plane fitting to locally smooth the matching cost values. We conduct experiments using the proposed method and compare it with belief propagation, census transform-based, and adaptive normalized cross correlation stereo matching methods. Experimental results show that our proposed matching method outperforms the test stereo matching methods under radiometric differences.

Abstract-In blocking-matching algorithms, a local window is used to measure the similarity (or dissimilarity) between pixels of a stereo pair. Although some area-based stereo matching methods have been developed and work well in many types of regions, such as textureless or object boundary regions, there are still cases where these methods are weak. In this paper, we propose an enhanced blocking-matching method to solve the correspondence problem in stereo matching. The proposed algorithm is an improved adaptive support weight method, which first classifies the pixels in the reference image and then uses the spatial weight variable window method and adaptive support weight method for each type of pixel. We also analyze and show the advantages and disadvantages of the variable window method and the adaptive support weight method and suggest solutions for cases where the two methods may not be ideal. The experimental results using the Middlebury images show that the proposed method outperform local stereo methods.Index Terms-Stereo matching, spatial weight, adaptive window.