Rank-Based Encoding Features for Stereo Matching

Abstract

We propose a novel feature extraction method and a cost function for stereo matching, that is robust and stable in matching images from different photographing conditions. Based on the Spearman rank correlation, for the pixel in the window centered around a certain pixel, the code is proposed as a rank sequence obtained from an ordered set of the pixel values. We also apply the same approach to X-gradient image and Y-gradient image. Finally, three rank sequences are obtained from the matching windows of original image, X-gradient image and Y-gradient image. They are used as the matching features of the center pixel. Then, the matching cost of two pixels will be defined as the weighted combination of the difference between their features. Our algorithm has a better matching effect than existing algorithms, especially in low-texture areas and object boundaries. We conducted experiments on Middlebury dataset that has different illumination and exposure images and KITTI dataset whose images were taken outdoor under radiometric distortions. The experimental results indicate that our algorithm is superior to the recently developed algorithms under radiometric variations, such as fuzzy encoding pattern1 and robust soft rank transform,3 whereas the speed is still fast.