Efficient and Accurate Stereo Matching via Multi-Scale Spatial Attention Fusion and Global Enhanced Error Perception

Depth estimation through stereo matching is a classic computer vision problem and has a wide range of practical applications. While recent advancements have demonstrated the potential of efficient stereo matching, most real-time stereo matching techniques encounter difficulties in predicting disparities around ill-posed regions and object details, leading to disparity maps that are relatively blurry and lacking precision in finer features. In this paper, we propose MG-Stereo, an accurate and efficient stereo matching framework that focuses on generating real-time and accurate disparity maps. We propose a Multi-Scale Spatial Attention fusion (MSAF) module to capture spatial information at different scales, more effectively combine context and geometry features, and improve the effect of cost aggregation. In addition, we propose a Global Enhanced Error Eerception (GEEP) module to further refine the initial disparity map by utilizing semantic information, high-frequency information, and attention guidance, effectively restoring ill-posed regions and object details that perform poorly in the initial disparity map. Our MG-Stereo achieves outstanding performance on public benchmarks and demonstrates strong generalization ability.