Locality-aware Gaussian Compression for Fast and High-quality Rendering

TL;DR

LocoGS is presented, a locality-aware 3D Gaussian Splatting framework that exploits the spatial coherence of 3D Gaussians for compact modeling of volumetric scenes and proposes a novel locality-aware 3D Gaussian representation that effectively encodes locally-coherent Gaussian attributes using a neural field representation with a minimal storage requirement.

Abstract

We present LocoGS, a locality-aware 3D Gaussian Splatting (3DGS) framework that exploits the spatial coherence of 3D Gaussians for compact modeling of volumetric scenes. To this end, we first analyze the local coherence of 3D Gaussian attributes, and propose a novel locality-aware 3D Gaussian representation that effectively encodes locally-coherent Gaussian attributes using a neural field representation with a minimal storage requirement. On top of the novel representation, LocoGS is carefully designed with additional components such as dense initialization, an adaptive spherical harmonics bandwidth scheme and different encoding schemes for different Gaussian attributes to maximize compression performance. Experimental results demonstrate that our approach outperforms the rendering quality of existing compact Gaussian representations for representative real-world 3D datasets while achieving from 54.6$\times$ to 96.6$\times$ compressed storage size and from 2.1$\times$ to 2.4$\times$ rendering speed than 3DGS. Even our approach also demonstrates an averaged 2.4$\times$ higher rendering speed than the state-of-the-art compression method with comparable compression performance.

Figures (15)

• Figure 1: Evaluation of the local coherence of Gaussian attributes. We visualize histograms of the Euclidean distances of Gaussian attributes (top), and bar graphs of the average Euclidean distances of Gaussian attributes (bottom) between two Gaussians at different spatial distances. The yellow histograms and bar graphs correspond to the largest spatial distances, while the pink ones correspond to the smallest spatial distances.
• Figure 2: Overview of our framework: locality-aware 3D Gaussian representation.
• Figure 3: Overall pipeline of LocoGS.
• Figure 4: Comparison with baseline approaches on Mip-NeRF 360 barron2022mip.
• Figure 5: Qualitative results on the 'stump' scene of Mip-NeRF 360 barron2022mip. Rendering speed, LPIPS, and storage size are shown at the bottom of each subfigure.