Parameter selection framework for stereo correspondence

Abstract

In this paper, we propose a method to select parameter values for stereo matching methods. The proposed method was trained in a supervised manner, and an evolutionary algorithm is used to select optimized parameter values for a given domain and a cost function constructed to measure the goodness level of candidate parameter values. Performance of the proposed method is compared to that of five current stereo matching methods, including the efficient large-scale stereo matching, belief propagation, semi-global block matching, stereo matching by training a convolutional neural network to compare image patches, and the efficient deep learning for stereo matching, for KITTI 2012, KITTI 2015, Middlebury, and EISAT datasets. The optimized parameters improve accuracy for all stereo matching methods considered, with some cases of improvement reaching up to $$24\%$$. Source code and experimental results are available online.