Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BOGausS: Better Optimized Gaussian Splatting

Stéphane Pateux, Matthieu Gendrin, Luce Morin, Théo Ladune, Xiaoran Jiang

TL;DR

This work targets efficient and high-fidelity novel-view synthesis via Gaussian Splatting by introducing a precision-aware optimization framework. It combines a variationally motivated precision model, an unbiased Sparse Adam optimizer, and density-preserving, distortion-aware densification to produce up to 10x lighter Gaussian sets without quality loss. Key contributions include a precision-aware update mechanism, inheritance of optimizer state after splitting, and pruning/splitting criteria grounded in rate-distortion-inspired metrics, yielding up to 0.5–1 dB PSNR gains over prior methods. The approach demonstrates strong quantitative and qualitative improvements across standard benchmarks, enabling high-quality renderings with reduced model complexity and improved robustness in varied camera setups.

Abstract

3D Gaussian Splatting (3DGS) proposes an efficient solution for novel view synthesis. Its framework provides fast and high-fidelity rendering. Although less complex than other solutions such as Neural Radiance Fields (NeRF), there are still some challenges building smaller models without sacrificing quality. In this study, we perform a careful analysis of 3DGS training process and propose a new optimization methodology. Our Better Optimized Gaussian Splatting (BOGausS) solution is able to generate models up to ten times lighter than the original 3DGS with no quality degradation, thus significantly boosting the performance of Gaussian Splatting compared to the state of the art.

BOGausS: Better Optimized Gaussian Splatting

TL;DR

This work targets efficient and high-fidelity novel-view synthesis via Gaussian Splatting by introducing a precision-aware optimization framework. It combines a variationally motivated precision model, an unbiased Sparse Adam optimizer, and density-preserving, distortion-aware densification to produce up to 10x lighter Gaussian sets without quality loss. Key contributions include a precision-aware update mechanism, inheritance of optimizer state after splitting, and pruning/splitting criteria grounded in rate-distortion-inspired metrics, yielding up to 0.5–1 dB PSNR gains over prior methods. The approach demonstrates strong quantitative and qualitative improvements across standard benchmarks, enabling high-quality renderings with reduced model complexity and improved robustness in varied camera setups.

Abstract

3D Gaussian Splatting (3DGS) proposes an efficient solution for novel view synthesis. Its framework provides fast and high-fidelity rendering. Although less complex than other solutions such as Neural Radiance Fields (NeRF), there are still some challenges building smaller models without sacrificing quality. In this study, we perform a careful analysis of 3DGS training process and propose a new optimization methodology. Our Better Optimized Gaussian Splatting (BOGausS) solution is able to generate models up to ten times lighter than the original 3DGS with no quality degradation, thus significantly boosting the performance of Gaussian Splatting compared to the state of the art.

Paper Structure

This paper contains 19 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. 3D Gaussian Splatting
  4. Densification Optimization
  5. Precision matrix derivation
  6. Improved gradient descent
  7. Proposed optimizer
  8. Precision-aware parameters updates
  9. Sparse and unbiased states updates
  10. States inheritance after Gaussian splitting
  11. Improved Densification
  12. Effective opacity-based Gaussian pruning
  13. Density-preserving Gaussian splitting
  14. Distortion-based prioritization of Gaussians splitting
  15. Experiments
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Qualitative visual comparison of our proposed method with MCMC approach MCMC24 on Garden scene with varying numbers of Gaussians. Our proposed method outperforms both visually and objectively MCMC thanks to a better reconstruction on background.
  • Figure 2: Comparative PSNR performances vs number of Gaussians. Unfilled-marks plots use exposure compensation while filled-marks plots don't.