Multi-view stereo 3D building reconstruction with sparse depth and edge location priors

Abstract

Accurate 3D building reconstruction remains challenging for large-scale structures with complex geometry. While deep learning-based Multi-View Stereo (MVS) methods improve upon traditional approaches, they exhibit errors in depth-discontinuous regions due to insufficient depth priors and architectural feature integration. To address these issues, this paper introduces ISENet, featuring: (1) an adaptive feature fusion framework for enhanced UAV image feature extraction, and (2) a multi-stage edge-aware depth hypothesis module leveraging sparse depth and edge location priors. Evaluations demonstrate state-of-the-art performance on the DTU dataset with Accuracy (0.238 mm) and Overall (0.277 mm) metrics. On real-world data, ISENet achieves 14.4 mm modeling error and 235 times higher point cloud density than commercial solutions. The approach also generalizes to standard MVS scenarios.