Learning stereo disparity with feature consistency and confidence

Most of the existing stereo matching methods have been formulated into four regular parts: feature extraction (FE), cost calculation (CC), cost aggregation (CA), and disparity refinement (DF). They can obtain high precision results in most regions through modifying parts of the four methods, but still have problems in some ill-posed regions. This paper focuses on feature consistency and confidence (FCC), discovers the new attributes of the feature, and proposes a novel neural network structure for stereo matching by measuring the consistency and confidence of features. Base on this method, the paper fuses the cost volume and calculates the pixel confidence map for cost calculation and cost aggregation. The experimental results show the proposed method outperforms most of the state-of-the-art methods on both SceneFlow and Kitti benchmarks and lowers the estimation error of stereo matching down to 1.82% ranking at the 7th position in the Kitti 2015 scoreboard six months ago (http://www.cvlibs.net/datasets/kitti/).