# Compact 3D Scene Representation via Self-Organizing Gaussian Grids

**Authors:** Wieland Morgenstern, Florian Barthel, Anna Hilsmann, Peter Eisert  
**arXiv:** [2312.13299](https://arxiv.org/abs/2312.13299)  
**Date:** 2023-12-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.13299v2/

## Key points

- The paper presents a novel approach to address the storage size limitations of 3D Gaussian Splatting (3DGS) for modeling static 3D scenes.
- Authors propose a compact scene representation method that organizes the parameters of 3DGS into a 2D grid with local homogeneity, leading to a drastic reduction in storage requirements without compromising visual quality during rendering.
- The method leverages perceptual redundancies present in natural scenes by demonstrating that multiple permutations of Gaussian parameters can equivalently represent a scene.
- A highly parallel algorithm is developed to arrange the high-dimensional Gaussian parameters into a 2D grid while preserving their neighborhood structure and enforcing local smoothness during training.
- The uncompressed Gaussians maintain the same structure as 3DGS, enabling seamless integration with established renderers.
- Authors achieve a remarkable reduction factor of 17x to 42x in size for complex scenes without increasing training time.
- Additional information on the project can be found on their project page at https://fraunhoferhhi.github.io/Self-Organizing-Gaussians/.
- Authors have added compression of spherical harmonics and updated compression methods using JPEG XL for all attributes, including qualitative comparisons of additional scenes and further advancements.

## Summary

The paper "Compact 3D Scene Representation via Self-Organizing Gaussian Grids" by Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert presents a novel approach to address the storage size limitations of 3D Gaussian Splatting (3DGS) for modeling static 3D scenes. While 3DGS offers fast rendering at high quality compared to Neural Radiance Fields, its storage requirements are significantly higher. This poses challenges for practical deployment on resource-constrained devices. To overcome this issue, the authors propose a compact scene representation method that organizes the parameters of 3DGS into a 2D grid with local homogeneity. This organization ensures a drastic reduction in storage requirements without compromising visual quality during rendering. Central to their idea is the explicit exploitation of perceptual redundancies present in natural scenes. By leveraging the inherent nature of scenes, they demonstrate that multiple permutations of Gaussian parameters can equivalently represent a scene. To achieve this, the authors develop a highly parallel algorithm that arranges the high-dimensional Gaussian parameters into a 2D grid while preserving their neighborhood structure. During training, they enforce local smoothness between the sorted parameters in the grid. The uncompressed Gaussians maintain the same structure as 3DGS, enabling seamless integration with established renderers. Through their method, the authors achieve a remarkable reduction factor of 17x to 42x in size for complex scenes without increasing training time. This advancement marks a significant leap forward in the domain of  and consumption. For additional information on their project, readers are directed to their project page at https://fraunhoferhhi.github.io/Self-Organizing-Gaussians/. Furthermore, the authors have added compression of spherical harmonics and updated compression methods with improved results using JPEG XL for all attributes. They include qualitative comparisons of additional scenes and move compression explanations and comparisons to the main paper. Additionally, they compare their work with "Making Gaussian Splats smaller," showcasing further advancements in .