Enhancing stereo matching with classification

Abstract

This paper presents a novel approach that employs classification to enhance the accuracy of the stereo matching problem. First, the images are treated in order to improve their pixel to pixel correspondence and reduce illumination differences. After that, stereo matching is addressed using different methods with emphasis on local ones like the sum of absolute distances and normalized cross correlation. Other state-of-the-art approaches are also considered. Then, and for every pixel, different features are computed from the input stereo image and the initially found depth map. Afterward, boosting and neural networks, as classification methods, are used to handle occlusion and enhance stereo matching by finding the erroneous disparity values. These values are then corrected through a completion stage. The accuracy of the proposed implementation improves on the problem in an efficient manner. A timing analysis of the method is provided to validate the real time performance. This paper further clarifies some of the possible developments based on various discussions.