DeferredGS: Decoupled and Relightable Gaussian Splatting With Deferred Shading.

Area of Science:

• Computer Graphics
• Computer Vision
• 3D Scene Reconstruction

Background:

• Neural Radiance Fields (NeRFs) offer realistic 3D reconstruction but are slow.
• Gaussian splatting accelerates rendering but struggles with independent texture/lighting editing.
• Existing methods for decoupling texture and lighting in Gaussian splatting face challenges with reflective scenes and relighting artifacts.

Purpose of the Study:

• To introduce DeferredGS, a novel method for decoupling and relighting Gaussian splatting representations.
• To enable independent editing of texture and lighting in 3D scenes.
• To achieve more realistic relighting effects and improve geometry plausibility.

Main Methods:

• DeferredGS models illumination with a learnable environment map.
• Additional attributes like texture parameters and normal direction are defined on Gaussians.
• Deferred shading is applied for enhanced relighting, with normals distilled from a signed distance function.

Main Results:

• DeferredGS successfully decouples texture and lighting in Gaussian splatting.
• The method produces plausible geometry and decomposition results, even on reflective scenes.
• Deferred shading leads to more realistic relighting compared to prior techniques.

Conclusions:

• DeferredGS significantly improves upon existing Gaussian splatting methods for 3D scene editing and relighting.
• The approach offers superior performance in novel view synthesis and relighting tasks.
• DeferredGS provides a robust solution for realistic 3D scene manipulation.