Monocular and Multi-View Geometry-Collaborated Multi-View Stereo Network

Learning-based multi-view stereo methods still struggle in ill-posed regions such as repetitive patterns, textureless and occluded areas, where photometric consistency fails to provide reliable matching cues. To address this challenge, a Monocular and multi-view Geometry-Collaborated multi-view stereo Network (MGCNet) is proposed that effectively and efficiently collaborates multi-view geometric cues with monocular priors from the vision foundation model for high-quality multi-view depth estimation. At the core of MGCNet is the proposed Monocular and Multi-view Geometry Coupled iterative Refinement (MGCR) module, which comprises two key components. First, a geometry-aware context encoder projects monocular priors into a latent feature space and fuses them with reference-view context features, enriching context representations with global geometric information while reducing sensitivity to monodepth inaccuracies. Second, a global geometry coupling unit enables efficient global cross-modal interaction between monocular and multi-view geometric cues for accurate depth residual learning. To further exploit the complementary strengths of cost filtering-based and iterative optimization-based refinement, an alternating depth refinement strategy is introduced, which interleaves cascaded cost filtering for enhancing multi-view geometric consistency with MGCR-based refinement targeting challenging regions. Extensive experiments on DTU, Tanks & Temples, and ETH3D benchmarks demonstrate that MGCNet achieves state-of-the-art performance with high efficiency, particularly excelling in challenging scenes with matching ambiguities.