RongLiu-Leo/beta-splatting

What it solves

Deformable Beta Splatting (DBS) addresses the limitations of 3D Gaussian Splatting (3DGS), specifically its struggle to capture complex geometries and diverse colors due to its reliance on Gaussian kernels and low-order Spherical Harmonics (SH). It aims to provide higher fidelity geometric details and better color representation while reducing memory usage and increasing rendering speed.

How it works

DBS replaces standard Gaussian kernels with deformable Beta Kernels, which feature bounded support and adaptive frequency control for more precise geometry. It also extends these Beta Kernels to color encoding to better represent diffuse and specular components compared to SH-based methods. Additionally, the project implements a distribution-preserved Markov chain Monte Carlo (MCMC) densification technique that relies on regularized opacity rather than specific kernel properties.

Who it’s for

This tool is designed for researchers and developers working on real-time radiance field rendering, 3D reconstruction, and computer vision.

Highlights

• Improved Efficiency: Uses only 45% of the parameters of 3DGS-MCMC while rendering 1.5x faster.
• Higher Fidelity: Captures fine geometric details and complex colors more effectively than Gaussian-based methods.
• Compact Storage: Includes a compression utility that can reduce storage requirements by 5x to 6x compared to standard PLY files.
• State-of-the-Art Quality: Achieves superior visual quality across various benchmarks including Mip-NeRF 360 and Tanks and Temples.