SVGS: Enhancing Gaussian Splatting Using Primitives With Spatially Varying Colors.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Gaussian Splatting offers impressive multi-view reconstruction using explicit Gaussian representations.
• Current methods use single view-dependent color and opacity per primitive, leading to inefficient scene representation.

Purpose of the Study:

• To introduce Spatially Varying Gaussian Splatting (SVGS) for improved representation ability.
• To enhance novel-view synthesis and geometric reconstruction quality.

Main Methods:

• SVGS utilizes spatially varying colors and opacity within single Gaussian primitives.
• Implemented spatially varying functions include bilinear interpolation, movable kernels, and tiny neural networks.
• Employs 2D Gaussian surfels as primitives.

Main Results:

• All three implemented functions outperformed the baseline Gaussian Splatting method.
• Movable kernels achieved superior novel view synthesis performance across multiple datasets.
• SVGS demonstrated enhanced representation ability for scenes with complex textures and simple geometry.

Conclusions:

• Spatially varying functions significantly improve Gaussian Splatting capabilities.
• SVGS shows strong potential for practical applications in real-world scene reconstruction.
• The proposed method enhances both visual fidelity and geometric accuracy in novel-view synthesis.