Symmetry Weight-sharing for Patch-based Stereo Matching

Abstract Neural networks are becoming more popular than traditional methods in stereo matching. The can be decomposed into four sub-modules: feature extraction / matching cost computation, aggregation, disparity computation optimization, and refinement. A typical design for the is that left right branches share same weights. However, Siamese weak at distinguishing neighboring patches because of interference geometric distortion on slanted surfaces. This paper proposes symmetry weight-sharing to improve networks. geometry patch comparison has been analyzed, which shows fulfill half-translation module proposed implement without additional computational costs. Experiments KITTI 2012 2015 datasets show have better performance