G2L-Stereo: Global to Local Two-Stage Real-Time Stereo Matching Network

Abstract

Developing fast and accurate stereo matching algorithms is crucial for real-world embedded vision applications. Depth information plays a significant role in scene understanding, and depth calculated through stereo matching is generally considered to be more precise and reliable than that obtained from monocular depth estimation. However, speed-oriented stereo matching methods often suffer from poor feature representation due to sparse sampling and detail loss caused by unreasonable disparity allocation during upsampling. To address these issues, we propose G2L-Stereo, a two-stage real-time stereo matching network that combines global disparity range prediction and local disparity range prediction. In the global disparity range prediction stage, we introduce feature-guided connections for cost aggregation, enhancing the expressive power of sparse features by aligning the feature space across different scales of cost volumes. We also incorporate confidence estimation into the upsampling algorithm to reduce the propagation of inaccurate disparities during upsampling, yielding more precise disparity maps. In the local disparity range prediction stage, we develop a disparity refinement module guided by neighborhood similarity. This module aggregates similar neighboring costs to estimate disparity residuals and refine disparities, restoring lost details in the low-resolution disparity map and further enhancing disparity accuracy. Extensive experiments on the SceneFlow and KITTI datasets validate the effectiveness of our model, showing that G2L-Stereo achieves fast inference while maintaining accuracy comparable to state-of-the-art methods.