Stereo Matching for Slanted Plane: A Probability Model

Many conventional stereo matching algorithms use frontal-parallel surface assumption. But in real world, there are many slanted surfaces which are not consistent with the assumption. So, errors are resulted from this assumption. In this paper, to handle slanted surfaces in stereo matching, a probability model is presented. The probability model gives a weight to measure the contributions of neighboring pixels to the probability of pixels assigned the ground truth. The weights are correlated to monocular information, namely, similarity of the color, spatial distance and binocular information, namely, dissimilarity property of pixels. Experimental results on test bed images with slanted plane and ground truth show that the proposed method produces accurate disparity maps.