Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CAT-3DGS Pro: A New Benchmark for Efficient 3DGS Compression

Yu-Ting Zhan, He-bi Yang, Cheng-Yuan Ho, Jui-Chiu Chiang, Wen-Hsiao Peng

TL;DR

CAT-3DGS Pro advances end-to-end compression for 3D Gaussian Splatting by replacing the triplane hyperprior with a PCA-guided vector-matrix (VM) hyperprior, enabling significantly reduced parameter count and faster decoding. It adds an alternate optimization strategy (A-RDO) and refines sampling-rate optimization (USRO) to improve the rate-distortion trade-off and training efficiency. The method achieves state-of-the-art rate-distortion performance across real-world scenes, with a reported 46.6% BD-rate reduction, ~3× faster training, and ~5× faster decoding on challenging scenes like Amsterdam. These improvements have practical implications for transmitting and storing high-quality 3DGS representations in real-time rendering and streaming applications. Future work envisions a unified compression framework for both position and attribute data within the ScaffoldGS representation to push RD performance even further.

Abstract

3D Gaussian Splatting (3DGS) has shown immense potential for novel view synthesis. However, achieving rate-distortion-optimized compression of 3DGS representations for transmission and/or storage applications remains a challenge. CAT-3DGS introduces a context-adaptive triplane hyperprior for end-to-end optimized compression, delivering state-of-the-art coding performance. Despite this, it requires prolonged training and decoding time. To address these limitations, we propose CAT-3DGS Pro, an enhanced version of CAT-3DGS that improves both compression performance and computational efficiency. First, we introduce a PCA-guided vector-matrix hyperprior, which replaces the triplane-based hyperprior to reduce redundant parameters. To achieve a more balanced rate-distortion trade-off and faster encoding, we propose an alternate optimization strategy (A-RDO). Additionally, we refine the sampling rate optimization method in CAT-3DGS, leading to significant improvements in rate-distortion performance. These enhancements result in a 46.6% BD-rate reduction and 3x speedup in training time on BungeeNeRF, while achieving 5x acceleration in decoding speed for the Amsterdam scene compared to CAT-3DGS.

CAT-3DGS Pro: A New Benchmark for Efficient 3DGS Compression

TL;DR

CAT-3DGS Pro advances end-to-end compression for 3D Gaussian Splatting by replacing the triplane hyperprior with a PCA-guided vector-matrix (VM) hyperprior, enabling significantly reduced parameter count and faster decoding. It adds an alternate optimization strategy (A-RDO) and refines sampling-rate optimization (USRO) to improve the rate-distortion trade-off and training efficiency. The method achieves state-of-the-art rate-distortion performance across real-world scenes, with a reported 46.6% BD-rate reduction, ~3× faster training, and ~5× faster decoding on challenging scenes like Amsterdam. These improvements have practical implications for transmitting and storing high-quality 3DGS representations in real-time rendering and streaming applications. Future work envisions a unified compression framework for both position and attribute data within the ScaffoldGS representation to push RD performance even further.

Abstract

3D Gaussian Splatting (3DGS) has shown immense potential for novel view synthesis. However, achieving rate-distortion-optimized compression of 3DGS representations for transmission and/or storage applications remains a challenge. CAT-3DGS introduces a context-adaptive triplane hyperprior for end-to-end optimized compression, delivering state-of-the-art coding performance. Despite this, it requires prolonged training and decoding time. To address these limitations, we propose CAT-3DGS Pro, an enhanced version of CAT-3DGS that improves both compression performance and computational efficiency. First, we introduce a PCA-guided vector-matrix hyperprior, which replaces the triplane-based hyperprior to reduce redundant parameters. To achieve a more balanced rate-distortion trade-off and faster encoding, we propose an alternate optimization strategy (A-RDO). Additionally, we refine the sampling rate optimization method in CAT-3DGS, leading to significant improvements in rate-distortion performance. These enhancements result in a 46.6% BD-rate reduction and 3x speedup in training time on BungeeNeRF, while achieving 5x acceleration in decoding speed for the Amsterdam scene compared to CAT-3DGS.

Paper Structure

This paper contains 15 sections, 5 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Proposed Method: CAT-3DGS Pro
  3. PCA-Guided Vector-Matrix (VM) Hyperprior
  4. Spatial Autoregressive Models (SARM) for Plane Coding
  5. Rate-distortion Optimization
  6. Experimental Results
  7. Implementation Details
  8. Rate-Distortion Comparison
  9. Complexity Comparison
  10. Training Time
  11. Decoding Time and Rendering Throughput
  12. Ablation Study
  13. Impact of Hyperprior and A-RDO on Compression Performance
  14. Sampling Rate Optimization
  15. Conclusion

Figures (4)

  • Figure 1: Comparison of rate-distortion performance, training time on the BungeeNeRF dataset, and decoding time for the Amsterdam scene.
  • Figure 2: System overview of our CAT-3DGS Pro. CARM: Channel-wise Autoregressive Models. SARM: Spatial Autoregressive Models.
  • Figure 3: Visualization of the bit-rate distribution on the Amsterdam scene for ViSRO and USRO. The bit rate indicates the number of bits needed to represent the anchor attributes of each anchor point.
  • Figure 4: Rate-distortion comparison of our CAT-3DGS Pro and several baseline methods.