GPR-MVS: Global Propagation Regularization for Large Scale Multi-view Stereo

Traditional learning-based Multi-view Stereo (MVS) relies on a decoupled Matching → Regularization pipeline, which leads to the loss of geometric information and fails to establish global consistency for large-scale scenes, particularly in challenging aerial scenarios. To address these issues, we propose the Global Propagation Regularization MVS network, named GPR-MVS, which achieves effective reconstruction in large-scale scenes by dynamically propagating inter-view global geometric constraints and performing depth-aware regularization. First, the proposed global dynamic propagation module constructs adaptive intra-view constraint propagation by coupling explicit similarity computation (cost volume construction) with implicit geometric optimization (regularization), while enabling cross-scale consistency alignment tailored for cascade architecture. Second, the proposed depth-aware regularization module introduces regularization mechanisms with varying hypothesis perception ranges: the coarse hypothesis stage constructs full sampling-range structures via epipolar attention, while the fine-grained stage employs a lightweight architecture for localized focus with reduced computational overhead. Additionally, a depth gradient-guided normalization layer is designed to suppress mismatched responses in edge regions with depth discontinuities embedding depth distribution priors. Qualitative and quantitative results on public datasets indicate that our method is effective and outperforms the SOTA.