Real-world 3D reconstruction often stumbles when a scene contains both shiny and matte surfaces, because existing methods tend to model these surface types separately. In a paper published in IEEE Transactions on Image Processing, Ye and colleagues propose MicroSDF, a neural implicit framework that unifies geometry and reflectance modeling through microfacet theory. The approach is built around a signed distance field (SDF) that simultaneously supports reflective and non-reflective surfaces.

The method introduces three main innovations. First, a microfacet-guided geometry model extracts both macroscopic and microfacet normals from the SDF, regularized by a microfacet normal consistency loss that encourages physically plausible surface orientations. Second, an enhanced dual-branch color model uses the microfacet normals for the specular branch, while the diffuse branch takes the reflection direction rather than the viewing direction, which the authors say better captures low-frequency components. Third, a detection-guided color blending strategy uses reflection priors to adaptively fuse the two color outputs, offering a more physically intuitive alternative to learned blending weights.

These components are combined with a multi-stage optimization scheme. Experiments on the DTU, Shiny Blender, Ref-NeRF, and DeepVoxels datasets report state-of-the-art performance, suggesting that microfacet-driven neural reconstruction is a promising direction for physically grounded 3D modeling. The paper is available on arXiv under the identifier 2610.19681.