Abstract Reconstructing accurate 3D geometry from images remains a core challenge in computer vision, especially in the presence of non-Lambertian reflectance and complex lighting conditions. This paper addresses the limitations of traditional multi-view photometric stereo techniquesabada2022using,abada2022improved, which often rely on separate pipelines for geometry and reflectance, leading to unstable results in shadowed or specular regions. We propose a novel neural architecture that unifies photometric and geometric cues in a single optimization framework. Our method comprises four neural sub-networks that jointly model surface occupancy, color appearance, spatially-changing surface reflectance, and specular basis decomposition. Unlike previous approaches, we eliminate explicit normal prediction and instead optimize them implicitly through a physically-based rendering equation that accounts for shadow visibility and surface reflectance. Shadowed regions are dynamically excluded via an online ray-based masking strategy, enhancing robustness under varying illumination. The framework is built upon a volume-to-surface rendering mechanism. Experimental evaluations on synthetic and real-world datasets demonstrate significant improvements in surface detail, normal consistency, and material separation compared to state-of-the-art multi-view photometric stereo methods. Our code will be available on github after acceptance at: https://github.com/lyabada/MVR-3D.
Gacem et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: