A New Split Algorithm for 3D Gaussian Splatting

Qiyuan Feng, Geng-Chen Cao, Haoxiang Chen, Tai-Jiang Mu, Ralph R. Martin, Shi-Min HuarXiv 2024 · DOI · arXiv

Introduction

3D Gaussian Splatting (3D GS) has emerged as a powerful technique for real-time neural rendering, offering an explicit and fast-to-render scene representation. However, the quality of this representation is often hampered by "messy" Gaussians—elements that are either excessively large or extremely elongated (needle-shaped). These issues lead to visual artifacts during scene editing and result in sparse, uneven point clouds. This paper presents a mathematically rigorous, closed-form algorithm designed to split these problematic Gaussians into smaller, uniform components while strictly preserving the original visual appearance.

The Challenge: Scale and Structural Inhomogeneity

While 3D GS provides high-fidelity rendering, it suffers from two primary forms of "inhomogeneity" that limit its utility for downstream geometric tasks:

  1. Scale Inhomogeneity: Large Gaussians can create "voids" or gaps, making it difficult to extract dense, high-quality point clouds from the model.
  2. Structural Inhomogeneity: Gaussians often become needle-like. During interactive editing, such as slicing an object, these elongated shapes produce blurred edges or protruding "spurs," ruining the precision of the edit.
Visualizing scale and structural inhomogeneity in 3D Gaussians and how the splitting algorithm resolves them.

Traditional methods for increasing Gaussian density typically involve simple copying and scaling, which fails to address these underlying shape deformities and can lead to training instability or visual artifacts during interactive manipulation.

The Proposed Splitting Algorithm

The authors introduce a novel splitting algorithm based on a 3D plane. The core principle is the conservation of visual appearance across the split. By treating the split as a conservation problem, the algorithm ensures that the resulting smaller Gaussians are visually indistinguishable from the original from any perspective. This is achieved by preserving three fundamental mathematical moments:

  • Zero-order moment: Conserves cumulative distribution of opacity.
  • First-order moment: Conserves the visual center (position).
  • Second-order moment: Conserves the mean of the outer product matrix (covariance/shape).

By formulating these conservation laws as integral tensor equations, the researchers derived a closed-form solution. This allows for the instantaneous calculation of new parameters (opacity, position, and anisotropic covariance) without the need for iterative optimization or additional training. The implementation also includes safeguards to maintain physical validity, ensuring covariance matrices remain positive definite despite floating-point precision constraints.

Experimental Evaluation

The algorithm was benchmarked using the standard NeRF Synthetic Dataset. The results demonstrate significant improvements across three key areas: interactive editing, training quality, and geometric reconstruction.

The proposed algorithm improves 3D Gaussian editing (left), training quality (center), and point cloud density (right).

Cleaner Editing and Rendering

In tasks involving curve removal and plane splitting, the method eliminated the "adhesion" artifacts typical of standard 3D GS. Quantitative metrics for "interval error" and "external excess" were significantly lower than baselines. Furthermore, integrating the split algorithm into the standard training loop improved rendering quality across the board:

  • PSNR: Improved from 36.93 to 37.19
  • SSIM: Increased from 0.9825 to 0.9839
  • LPIPS (Error): Decreased from 0.0236 to 0.0216

Enhanced Point Cloud Extraction

The algorithm effectively fills gaps in untextured regions, producing point clouds that are significantly more uniform. This makes the resulting 3D GS models much more suitable for traditional geometry processing tasks, such as normal estimation and mesh conversion.

Conclusion and Future Work

By providing a mathematically grounded way to subdivide Gaussians, this work bridges the gap between high-speed neural rendering and precise geometric editing. While the current focus is on splitting, future research may explore the inverse process—merging Gaussians—to enable better model compression. Additionally, extending the planar split logic to handle complex curved surfaces directly could further enhance the precision of geometric modifications in 3D Gaussian Splatting.

Summary updated

Abstract

3D Gaussian splatting models, as a novel explicit 3D representation, have been applied in many domains recently, such as explicit geometric editing and geometry generation. Progress has been rapid. However, due to their mixed scales and cluttered shapes, 3D Gaussian splatting models can produce a blurred or needle-like effect near the surface. At the same time, 3D Gaussian splatting models tend to flatten large untextured regions, yielding a very sparse point cloud. These problems are caused by the non-uniform nature of 3D Gaussian splatting models, so in this paper, we propose a new 3D Gaussian splitting algorithm, which can produce a more uniform and surface-bounded 3D Gaussian splatting model. Our algorithm splits an $N$-dimensional Gaussian into two N-dimensional Gaussians. It ensures consistency of mathematical characteristics and similarity of appearance, allowing resulting 3D Gaussian splatting models to be more uniform and a better fit to the underlying surface, and thus more suitable for explicit editing, point cloud extraction and other tasks. Meanwhile, our 3D Gaussian splitting approach has a very simple closed-form solution, making it readily applicable to any 3D Gaussian model.
BibTeX
@misc{feng2024new,
  author = {Qiyuan Feng and Geng-Chen Cao and Haoxiang Chen and Tai-Jiang Mu and Ralph R. Martin and Shi-Min Hu},
  title = {{A New Split Algorithm for 3D Gaussian Splatting}},
  year = {2024},
  eprint = {2403.09143},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.09143},
}
Figure