Segmentation-based stereo matching using combinatorial similarity measurement and adaptive support region

Abstract

An accurate segmentation-based algorithm using combinatorial similarity measurement and adaptive support aggregation strategy is developed for stereo matching. The advantage of our method which based on a segmentation framework is generating disparity map in textureless regions correctly and localize depth boundaries precisely. Initial disparity map is estimated by a local correspondence approach which consisting of a combinatorial similarity measurement function and an adaptive window with arbitrary shape and size. The accuracy of the initial disparity map generated by combinatorial similarity measurement is better than that generated by individual similarity measurement. The plane parameters are fitted utilizing the initial disparity map. In this process, a method named “MC-RANSAC” which consisting of mutual consistency check criterion and random sample consensus algorithm is proposed to filter out the outliers and obtain the stable disparity plane parameters. Lastly, the disparity plane is optimized by belief propagation. Experiments with stereo image pairs show the validity of the proposed method.