MesonGS++: Post-training Compression of 3D Gaussian Splatting with Hyperparameter Searching
Summary
MesonGS++ is a novel post-training compression codec designed to significantly reduce the storage cost of 3D Gaussian Splatting (3DGS) models, which are known for high-quality novel view synthesis but suffer from prohibitive storage requirements. Unlike prior methods that struggle with numerous coupled hyperparameters, MesonGS++ integrates joint importance-based pruning, octree geometry coding, attribute transformation, selective vector quantization for higher-degree spherical harmonics, and group-wise mixed-precision quantization with entropy coding. It also features a configuration side that optimizes reserve ratio and bit-width allocation using discrete sampling and 0-1 integer linear programming to meet specific storage budgets. A linear size estimator and a CUDA parallel quantization operator accelerate the hyperparameter search. Experiments demonstrate MesonGS++ achieves over 34x compression while maintaining rendering fidelity, outperforming existing methods and accurately hitting target sizes. It can even exceed vanilla 3DGS PSNR at 20x compression on the Stump scene without retraining.
Key takeaway
For research scientists developing or deploying 3D Gaussian Splatting applications, you should evaluate MesonGS++ to drastically reduce model storage requirements while preserving or even enhancing rendering quality. Its ability to accurately meet target size budgets and achieve over 34x compression without retraining offers a significant advantage for practical deployment on resource-constrained platforms, making high-fidelity 3DGS more viable.
Key insights
MesonGS++ compresses 3DGS models over 34x by optimizing hyperparameter search for target storage budgets.
Principles
- Joint optimization of compression parameters improves rate-distortion.
- Post-training methods can surpass uncompressed baselines.
- Hyperparameter search benefits from accelerated estimation.
Method
MesonGS++ combines importance-based pruning, octree geometry coding, attribute transformation, selective vector quantization, and group-wise mixed-precision quantization, optimizing reserve ratio and bit-width via discrete sampling and 0-1 integer linear programming.
In practice
- Achieve 34x 3DGS compression with MesonGS++.
- Surpass vanilla 3DGS PSNR at 20x compression.
- Use CUDA for faster quantization operations.
Topics
- 3D Gaussian Splatting
- Post-training Compression
- Hyperparameter Optimization
- Rate-Distortion Optimization
- Novel View Synthesis
Code references
Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.