Patch based confidence prediction for dense disparity map

Abstract

In this paper, we propose a novel method to predict the correctness of stereo correspondences, which we call confidence, and a confidence fusion method for dense disparity estimation. The input of our method consists in a two channels local window (disparity patch) which is designed by taking into account ideas of conventional confidence features. 1st channel is coming from the idea that neighboring pixels which have consistent disparities are more likely to be correct matching. In 2nd channel, a disparity from another image is considered such that the matches from left to right image should be consistent with those from right to left. The disparity patches are used as inputs of Convolutional Neural Networks so that the features and classifiers are simultaneously trained unlike what is done by existing methods. Moreover, the confidence is incorporated into Semi-Global Matching(SGM) by adjusting its parameters directly. We show the prominent performance of both confidence prediction and dense disparity estimation on KITTI datasets which are real world scenery.