Jaccard distance as similarity measure for disparity map estimation

Abstract

High confidence in disparity map estimation is critical in several application fields. A novel framework that employs customized local binary patterns and Jaccard distance for stereo matching along stereo consistency checks is presented. The proposal contributes with a method that allows greater confidence in its estimates, without dependence on supervised learning, and capable of generating a dense map with low-cost filtering. The proposed framework has been implemented in CPU and GPU for parallel processing capability. First, Local binary patterns are obtained during the initial stage; then, the Jaccard distance is employed as a similarity measure in the stereo matching stage; subsequently, a matching consistency check is performed, and singular disparities are removed. A comparison among novel and state-of-the-art algorithms for sparse disparity map estimation is performed employing Middlebury and KITTI stereo Datasets where the quality criteria used were percentage of bad pixels (B), quantity of invalid pixels, processing time and running environments to put each framework into context, obtaining down to 2.07% bad matching pixels and performing better than state-of-the-art cost functions