Performance evaluation of deep learning dense matching models for high-resolution satellite image-based 3D reconstruction

Accurate dense matching is fundamental to high-precision 3D reconstruction from sub-meter satellite imagery. Traditional methods (e.g., SGM) often struggle in weakly-textured or occluded regions, whereas deep learning methods demonstrate both high accuracy and strong robustness in such challenging areas. Therefore, deep learning-based dense matching algorithms combined with high-resolution satellite imagery have broad prospects in fields such as smart cities and disaster monitoring. In this study, we benchmark five deep learning models (IGEV-Stereo, PSM-Net, HMSM-Net, Stereo-Net and S3Net) on the Urban Semantic 3D (US3D) dataset, covering three typical scene types: dense urban environments, complex natural terrain and roads/bridges. The results show that HMSM-Net and S3Net can perform high-precision reconstruction of complex scenes such as urban buildings, enabling city-level 3D visualization and monitoring of urban operations. Meanwhile, IGEV-Stereo has the shortest inference time and provides a good balance between accuracy and speed, making it particularly suitable for time-sensitive applications such as disaster monitoring.