TY - JOUR
T1 - GPR-MVS
T2 - Global Propagation Regularization for Large-Scale Multiview Stereo
AU - Zhang, Wei
AU - Wu, Yihang
AU - Yu, Shengkai
AU - Li, Songhua
AU - Li, Qiang
AU - Wang, Qi
N1 - Publisher Copyright:
© 1980-2012 IEEE.
PY - 2026
Y1 - 2026
N2 - Traditional learning-based multiview stereo (MVS) relies on a decoupled matching \to 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 (GPR-MVS) network, which achieves effective reconstruction in large-scale scenes by dynamically propagating interview global geometric constraints and performing depth-aware regularization. First, the proposed global dynamic propagation module (GDPM) constructs adaptive intraview 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 (DARM) 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. In addition, a depth gradient-guided normalization layer (DGGNL) 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.
AB - Traditional learning-based multiview stereo (MVS) relies on a decoupled matching \to 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 (GPR-MVS) network, which achieves effective reconstruction in large-scale scenes by dynamically propagating interview global geometric constraints and performing depth-aware regularization. First, the proposed global dynamic propagation module (GDPM) constructs adaptive intraview 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 (DARM) 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. In addition, a depth gradient-guided normalization layer (DGGNL) 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.
KW - 3-D reconstruction
KW - dense image matching
KW - multiview stereo (MVS)
UR - https://www.scopus.com/pages/publications/105043922371
U2 - 10.1109/TGRS.2026.3710063
DO - 10.1109/TGRS.2026.3710063
M3 - 文章
AN - SCOPUS:105043922371
SN - 0196-2892
VL - 64
JO - IEEE Transactions on Geoscience and Remote Sensing
JF - IEEE Transactions on Geoscience and Remote Sensing
M1 - 5634710
ER -