2DGH: 2D Gaussian-Hermite Splatting for High-Quality Rendering and Better Geometry Features

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• 2D Gaussian Splatting is a key method for simultaneous 3D reconstruction and novel view synthesis.
• Traditional Gaussian kernels in splatting methods lack anisotropy and deformation, causing blurred object edges and limiting reconstruction quality.

Purpose of the Study:

• To enhance the representational power of Gaussian Splatting primitives.
• To improve the quality of 3D reconstruction and novel view synthesis.

Main Methods:

• Proposed using Gaussian-Hermite kernels as novel primitives within the Gaussian Splatting framework.
• Developed a unified mathematical formulation extending the standard Gaussian function.

Main Results:

• Gaussian-Hermite kernels demonstrated superior performance compared to traditional Gaussian kernels.
• Achieved more accurate geometry reconstruction on the DTU dataset.
• Showcased improved novel view synthesis on MipNeRF360 and a custom Detail dataset.

Conclusions:

• The Gaussian-Hermite kernel significantly enhances 3D reconstruction and novel view synthesis capabilities.
• This new primitive offers potential for high-quality 3D reconstruction and rendering applications.