GlossyGS: Inverse Rendering of Glossy Objects With 3D Gaussian Splatting

Area of Science:

• Computer Graphics
• Computer Vision
• 3D Reconstruction

Background:

• Neural Radiance Fields (NeRF) offer impressive reconstruction but are slow.
• 3D Gaussian Splatting (3D-GS) provides faster inverse rendering but struggles with glossy object details.
• Inverse rendering ambiguities hinder accurate geometry and material capture for reflective surfaces.

Purpose of the Study:

• To develop an advanced 3D-GS framework for high-fidelity reconstruction of glossy objects.
• To address the limitations of existing methods in capturing realistic geometry and materials for reflective surfaces.
• To introduce a novel approach that integrates material priors into 3D-GS for improved inverse rendering.

Main Methods:

• Introduced GlossyGS, a 3D-GS-based inverse rendering framework.
• Integrated micro-facet geometry segmentation priors to reduce ambiguities.
• Employed a normal map prefiltering strategy for accurate normal distribution simulation.
• Utilized a hybrid explicit-implicit representation for glossy object depiction.

Main Results:

• Achieved precise reconstruction of glossy object geometry and materials.
• Demonstrated significant improvements in handling reflective surface properties.
• Quantitative and qualitative analyses confirmed high-fidelity results.
• Outperformed existing state-of-the-art methods in glossy object reconstruction.

Conclusions:

• GlossyGS effectively reconstructs high-fidelity geometry and materials for glossy objects.
• The integration of material priors and hybrid representations overcomes key inverse rendering challenges.
• The proposed method offers a significant advancement in 3D reconstruction of reflective surfaces.