A relative depth-guided multiview stereo reconstruction framework for remote sensing imagery

The use of satellite imagery for automated 3D reconstruction plays a crucial role in various fields and has emerged as a key area of research at the intersection of photogrammetry and computer vision. The recent deep learning techniques have significantly enhanced stereo matching and feature representation, driving progress in this field. The recent deep learning techniques have significantly enhanced multi-view stereo (MVS), driving progress in this field. However, the current methods for constructing the cost volume of remote sensing images through deep learning lack effective information transfer and correlation and have high computational complexity. This paper proposes an efficient 3Dreconstruction method for satellite imagery, which utilizes a dynamic MVS framework for cross-scale information propagation and enables rapid height inference. The proposed method incorporates relative depth maps as priors into the multi-scale feature representations within the framework, thereby enabling more accurate extraction of geometric structures. The experiments validated the effectiveness of the method on WHU-TLC datasets, outperforming existing methods by significantly improving computational efficiency while maintaining accuracy.