CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression

ICLR 2025
National Yang Ming Chiao Tung University, Taiwan
National Chung Cheng University, Taiwan

*Indicates Equal Contribution
MY ALT TEXT

We aim to exploit much of the inter correlation between the Gaussian primitives by using triplane-based hyperprior and also achieve rate-distortion performance on several commonly used real-world datasets. (CARM: Channel- wise Autoregressive Models. SARM: Spatial Autoregressive Models.)

Abstract

3D Gaussian Splatting (3DGS) has recently emerged as a promising 3D representation. However, its substantial storage requirements necessitate efficient compression for transmission and application. This work introduces a novel rate-distortion-optimized compression framework for 3DGS, termed CAT-3DGS. We apply masking and coding techniques within ScaffoldGS to ensure efficient data transmission. For coding, we utilize the triplane hyperprior and employ channel-wise autoregressive models to predict probabilities for entropy coding. Moreover, an improved masking mechanism further enhances efficieny.

Combined with these features, CAT-3DGS achieves state-of-the-art compression performance on real-world datasets. On the Mip-NeRF 360 dataset, our CAT-3DGS achieves (at its second highest rate point) 78× and 26x rate reductions than 3DGS and ScaffoldGS, respectively, while achieving slightly higher PSNR by 0.16 dB.

Main Method

Our work, CAT-3DGS, introduces a rate-distortion-optimized approach that leverages context-adaptive triplanes to improve compression efficiency. By aligning multi-scale triplanes with the principal axes of Gaussian primitives, we capture spatial correlations for spatial autoregressive coding (e), while channel-wise autoregressive coding (d) exploits intra dependencies within each individual Gaussian primitive. A view frequency-aware masking mechanism further refines the process by skipping primitives with minimal impact on rendering quality.

MY ALT TEXT

Overview of our CAT-3DGS framework.

Results

MY ALT TEXT MY ALT TEXT

Visual Comparison

Before 1 After 1
Scaffold-GS (24.5dB/248.0MB)
Ours (25.0dB/21.4MB)
Before 2 After 2
HAC (27.3dB/31.8MB)
Ours (27.6dB/32.0MB)

BibTeX

@inproceedings{zhan2025cat3dgs,
        author    = {Yu-Ting Zhan and Cheng-Yuan Ho and Hebi Yang and Yi-Hsin Chen and Jui Chiu Chiang and Yu-Lun Liu and Wen-Hsiao Peng},
        title     = {{CAT-3DGS: A context-adaptive triplane approach to rate-distortion-optimized 3DGS compression}},
        booktitle = {Proceedings of the Thirteenth International Conference on Learning Representations (ICLR)},
        year      = {2025},
      }