SRNet: Self-supervised structure regularization for stereo matching

Jun Cheng

doi:10.1016/j.neucom.2025.131907

SRNet: Self-supervised structure regularization for stereo matching

2025-10-24

0

PRE

AI

DOI：10.1016/j.neucom.2025.131907

Original

Original request for help

Save

Abstract

En 中文

• We propose a novel framework to regularize the stereo matching using low-level structure constraints. The results show that the performance of many existing data-driven stereo matching networks can be improved by the proposed regularization. • We also propose a unique disparity aggregation module to leverage the association between the disparity and RGB information in the low-level structure detection. • We explore the use of the edge and keypoint as self-supervised labels for regularization and find that these different regularization terms work comparably to manual ground truth structure in the Scene Flow dataset. • We integrate the proposed regularization term with four different stereo matching networks, experimental results show that it is able to improve the performance and is generic for different networks.

SRNet: Self-supervised structure regularization for stereo matching

Abstract

Journal

Organization