Locality-aware Gaussian Compression for Fast and High-quality Rendering
Seungjoo Shin, Jaesik Park, Sunghyun Cho

TL;DR
LocoGS introduces a locality-aware Gaussian compression framework that significantly reduces storage and accelerates rendering of 3D scenes while maintaining high quality, outperforming existing methods.
Contribution
The paper proposes a novel locality-aware 3D Gaussian representation and a comprehensive compression framework that enhances efficiency and quality in 3D scene rendering.
Findings
Achieves 54.6× to 96.6× compression compared to existing methods.
Provides 2.1× to 2.4× faster rendering speeds.
Outperforms state-of-the-art compression methods in quality and speed.
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…
Peer Reviews
Decision·ICLR 2025 Poster
1. The overall solution is comprehensive, incorporating careful designs for initialization, pruning, and compression schemes for different components. 2. The paper is well-written with a clear logical flow. 3. The dense initialization demonstrates an interesting improvement in compression. 4. The analysis of storage size in Tab.3 provides a valuable indication for the community about the current bottlenecks in compression.
1. Number of Gaussians in Variants: It would be beneficial to provide the number of Gaussians for different variants, as Ours-Sparse has a lower FPS compared to Ours-Small and Ours. Understanding whether the final number of Gaussians influenced the results is important. Including the point number counts along the training iteration would help illustrate the influence of different initializations. Additionally, it would be interesting to see if such an initialization design could improve the perf
The concept of locality is not novel in 3D Gaussian Splatting, but the paper effectively illustrates this through graphs. Moreover, it clearly addresses the differences between other locality-based methods (anchor-based methods), explaining why the proposed method is important and highlighting key points for readers already familiar with anchor-based representation. The performance improvements in terms of size, rendering speed, and rendering quality are significant. Additionally, the paper in
As I mentioned in the strength section, the idea of utilizing locality in 3D Gaussian Splatting is not novel.
1. LocoGS carefully analyzes the relationships among 3D Gaussian attributes and introduces a locality-aware representation based on these relationships. 2. The paper is well written and easy to follow, with an excellent categorization of compression methods. 3. LocoGS outperforms existing methods in both compression performance and rendering speed.
1. Dense initialization is not related to compression; it is merely a warm-up trick. It would be unfair for the authors to use this trick since all of the baselines still initialize with COLMAP points. 2. As we know, the training time for Nerfacto is significantly longer than for 3DGS. Why do the authors choose to use Nerfacto for warm-up instead of 3DGS? Both methods can generate coarse depth maps, which can then be used to create an initialization point cloud. Additionally, how many iterations
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsComputer Graphics and Visualization Techniques · 3D Shape Modeling and Analysis · Video Analysis and Summarization
MethodsSPEED: Separable Pyramidal Pooling EncodEr-Decoder for Real-Time Monocular Depth Estimation on Low-Resource Settings
