Artistic gradient-based stereo matching algorithm with a hybrid local-global reasoning

Abstract

Dense stereo matching is one of the challenging problems in computer vision. It forms the basis for the extraction of three-dimensional scene structure through getting a dense disparity map. In this work, a new segment-based stereo matching algorithm with a gradient similarity and a global energy minimization technique is proposed (GcKuwa_Grad). It has two main stages: a local stage which aims at inferring all valid planes in disparity space using RANSAC and SVD and producing a good initialization for the global optimization space which aims at assigning memberships to these planes to all pixels in the reference image. The local stage adopts a new effective edge preserving Artistic Kuwahara cost volume filtering forming an important base for a new data term employed globally in segment domain for more reliable disparity plane assignment. GcKuwa_Grad represents a strong candidate with the top ranked methods according to Middlebury benchmark and with a first rank as a global gradient-based stereo matching algorithm.