# Compact 3D Gaussian Representation for Radiance Field

**Authors:** Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, Eunbyung Park  
**arXiv:** [2311.13681](https://arxiv.org/abs/2311.13681)  
**Date:** 2023-11-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.13681v1/

## Key points

- Neural Radiance Fields (NeRFs) capture complex 3D scenes with high fidelity
- Computational bottleneck caused by volumetric rendering hinders widespread adoption of NeRFs
- 3D Gaussian splatting (3DGS) offers faster rendering speed and promising image quality
- Study focuses on reducing number of Gaussian points without sacrificing performance and compressing Gaussian attributes
- Learnable mask strategy reduces number of Gaussians while preserving high performance
- Compact representation of view-dependent color using grid-based neural field instead of spherical harmonics
- Codebooks learned for compact representation of geometric attributes through vector quantization
- Approach achieves over 10 times reduced storage requirements and enhanced rendering speed compared to 3DGS
- Comprehensive framework for 3D scene representation combining high performance, fast training, compactness, and real-time rendering capabilities

## Summary

Neural Radiance Fields (NeRFs) have shown great promise in capturing complex 3D scenes with high fidelity. However, their widespread adoption is hindered by the computational bottleneck caused by volumetric rendering. In contrast, 3D Gaussian splatting (3DGS) offers a faster rendering speed and promising image quality by leveraging a 3D Gaussian-based representation and adopting the rasterization pipeline instead of volumetric rendering. To address this issue, this study focuses on two key objectives: reducing the number of Gaussian points without sacrificing performance and compressing the Gaussian attributes such as view-dependent color and covariance. The authors propose a learnable mask strategy that significantly reduces the number of Gaussians while preserving high performance. Additionally, they introduce a compact yet effective representation of view-dependent color using a grid-based neural field instead of relying on spherical harmonics. Furthermore, codebooks are learned to compactly represent the geometric attributes of Gaussians through vector quantization. Extensive experiments demonstrate that compared to 3DGS, this approach achieves over 10 times reduced storage requirements and enhanced rendering speed while maintaining scene representation quality. The work provides a comprehensive framework for 3D scene representation that combines high performance, fast training, compactness, and real-time rendering capabilities. The authors' project page can be found at https://maincold2.github.io/c3dgs/.