MicroSDF: Microfacet-Driven Hybrid Neural SDFs for Mixed-Reflectance Surface Reconstruction

Area of Science:

• Computer Vision
• Computer Graphics
• Neural Rendering

Background:

• Accurate 3D reconstruction is challenging due to diverse surface properties like reflectivity.
• Existing methods often struggle with combined reflective and non-reflective surfaces, limiting generalizability.
• A unified approach is needed for physically plausible 3D scene modeling.

Purpose of the Study:

• To introduce MicroSDF, a novel neural implicit framework for unified 3D reconstruction.
• To enable robust modeling of both geometry and reflectance across various surface types.
• To enhance physical plausibility in neural reconstruction through microfacet theory.

Main Methods:

• Developed a microfacet-guided geometry model extracting multi-scale surface normals from a signed distance field (SDF).
• Implemented an enhanced dual-branch color model using microfacet normals for high-frequency reflectance and reflection direction for diffuse components.
• Introduced a detection-guided color blending strategy for adaptive fusion of color outputs based on reflection priors.

Main Results:

• Achieved robust and high-fidelity 3D reconstruction across diverse datasets (DTU, Shiny Blender, Ref-NeRF, DeepVoxels).
• Demonstrated state-of-the-art performance compared to existing 3D reconstruction methods.
• Validated the physical plausibility and effectiveness of the proposed microfacet-based approach.

Conclusions:

• MicroSDF offers a unified and physically grounded solution for 3D reconstruction challenges posed by mixed-reflectivity surfaces.
• The framework establishes a new direction for high-fidelity neural reconstruction by integrating microfacet theory.
• MicroSDF significantly advances the capabilities of neural implicit representations for real-world scene modeling.