NormalMVS: Learning Efficient Geometry-Aware Multiview Stereo With Surface Normal Priors

Learning-based multiview stereo (MVS) has achieved remarkable progress, largely driven by cascade-based architectures. However, existing methods inevitably suffer from error accumulation across stages because of simple upsampling. Furthermore, relying solely on RGB images often leads to photometric ambiguity in textureless or non-Lambertian regions and a lack of explicit 3D constraints, resulting in erroneous correspondences. To address these challenges, we propose NormalMVS, an efficient framework that explicitly incorporates surface normal priors to enhance reconstruction quality. Specifically, we design a geometry-aware fusion (GAF) module that synergistically combines RGB and normal features and extracts discriminative geometric cues to resolve photometric ambiguities. To mitigate error propagation, we design a normal-guided depth refinement (NGDR) module. Unlike conventional bilinear interpolation, NGDR refines upsampled depth maps on the basis of local coplanarity and is supervised by a novel neighbor-weighted consistency loss that enforces geometric coherence. Additionally, a four-stage architecture and a sparse sampling strategy are introduced to significantly reduce the computational and memory overhead of 3D CNNs. Extensive experiments on the DTU, Tanks & Temples, and ETH3D benchmarks demonstrate that NormalMVS outperforms most existing methods in terms of efficiency, effectively balancing reconstruction quality and resource consumption.