TC-GS: Tri-plane based compression for 3D Gaussian Splatting

2025-03-26

AI Summary

To address the low compression efficiency and high memory overhead of 3D Gaussian Splatting (3DGS) caused by disordered Gaussian parameter distributions, this paper proposes a structured compression framework based on triplane representation. Our method maps unstructured Gaussian attributes—namely, positions, covariances, opacities, and spherical harmonic coefficients—onto a canonical triplane grid, enabling spatially coherent encoding. We further design a position-aware KNN decoder that explicitly models geometric and attribute correlations among neighboring Gaussians. Additionally, we introduce an adaptive wavelet loss to improve high-frequency detail reconstruction fidelity. Evaluated on multiple standard benchmarks, our approach significantly outperforms existing 3DGS compression methods: it reduces GPU memory consumption by 38%–52% while maintaining state-of-the-art novel-view synthesis quality.

Technology Category

• Computer Vision: 3D Computer Vision
• Data Mining & Knowledge Management: Data Compression
• Machine Learning: Deep Generative Models & Autoencoders

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Recently, 3D Gaussian Splatting (3DGS) has emerged as a prominent framework for novel view synthesis, providing high fidelity and rapid rendering speed. However, the substantial data volume of 3DGS and its attributes impede its practical utility, requiring compression techniques for reducing memory cost. Nevertheless, the unorganized shape of 3DGS leads to difficulties in compression. To formulate unstructured attributes into normative distribution, we propose a well-structured tri-plane to encode Gaussian attributes, leveraging the distribution of attributes for compression. To exploit the correlations among adjacent Gaussians, K-Nearest Neighbors (KNN) is used when decoding Gaussian distribution from the Tri-plane. We also introduce Gaussian position information as a prior of the position-sensitive decoder. Additionally, we incorporate an adaptive wavelet loss, aiming to focus on the high-frequency details as iterations increase. Our approach has achieved results that are comparable to or surpass that of SOTA 3D Gaussians Splatting compression work in extensive experiments across multiple datasets. The codes are released at https://github.com/timwang2001/TC-GS.

Problem

• Compress 3D Gaussian Splatting data to reduce memory cost
• Structure unstructured attributes using tri-plane encoding
• Enhance compression by leveraging Gaussian correlations and position priors

Innovation

• Tri-plane encoding for Gaussian attributes compression
• KNN decoding for Gaussian distribution correlation
• Adaptive wavelet loss for high-frequency details