Improving Disparity Consistency With Self-Refined Cost Volumes for Deep Learning-Based Satellite Stereo Matching

Abstract

Stereo matching algorithms are considered one of the most important subtasks in 3-D reconstruction, as 3-D coordinates are derived from the disparity values of pixels obtained through stereo matching. Recently, deep learning-based satellite stereo matching algorithms have been widely investigated, as they can capture both deep and shallow features of complex satellite scenes. However, several problems in satellite stereo matching, due to the unique properties of satellite images, remain unsolved, particularly in textureless and repetitive regions. In these regions, a single object in a satellite image is likely to be matched with similar objects, causing multiple disparity probabilities and shifts in the disparity estimation. To address the problem of disparity shifts, we propose a novel cost volume refinement strategy (CVRS). CVRS introduces both left-right and left-left cost volumes, which work together to refine disparities and eliminate false matches in textureless or repetitive regions, while preserving the original disparity values. With CVRS, we propose a new model for satellite stereo matching, the self-refined cost volume network (SRCV-Net). We evaluated CVRS and SRCV-Net on the US3D and WHU-Stereo datasets, comparing it using the EPE and D1 metrics. The application of CVRS demonstrated performance improvements in all models, and SRCV-Net achieved superior accuracy in satellite stereo matching. Furthermore, CVRS can be easily applied to various models with minimal structural changes and a small increase in parameters. SRCV-Net, with its innovative CVRS, provides an effective solution to the challenges of satellite stereo matching, offering enhanced accuracy, efficiency, and adaptability.