SGM-based Refined Hints for Stereo Matching Generalization Improvement leveraging Gaussian Policy

Stereo matching has been a hot research topic in 3D reconstruction. Although great progress has been made thanks to deep learning technique, the performance still suffers a significant drop when dealing with a new domain. Recently, LiDAR and Radar guidance strategies are explored to alleviate the generalization problem. However, they rely on extra depth sensor cues. In this paper, a novel SGM-based refined hints guided stereo matching paradigm is developed, which does not require corresponding depth sensor data. The proposed method includes disparity hints acquisition and disparity hints guided stereo matching. The former utilizes a divide-and-conquer technique to detect, replace, and refine SGM-based disparity to ensure the uniformity and density of hints, while the latter leverages the Gaussian policy to incorporate the location information of hints to enhance the effective matching, and thus improve the out-of-domain performance. Extensive experiments show a significant generalization improvement ($e.g$., from 4.78 to 1.03 average errors on the KITTI 2015 dataset).