SizeGS: Size-aware Compression of 3D Gaussians with Hierarchical Mixed Precision Quantization

Shuzhao Xie,Jiahang Liu,Weixiang Zhang,Shijia Ge,Sicheng Pan,Chen Tang,Yunpeng Bai,Zhi Wang

2024-12-08

Abstract:Effective compression technology is crucial for 3DGS to adapt to varying storage and transmission conditions. However, existing methods fail to address size constraints while maintaining optimal quality. In this paper, we introduce SizeGS, a framework that compresses 3DGS within a specified size budget while optimizing visual quality. We start with a size estimator to establish a clear relationship between file size and hyperparameters. Leveraging this estimator, we incorporate mixed precision quantization (MPQ) into 3DGS attributes, structuring MPQ in two hierarchical level -- inter-attribute and intra-attribute -- to optimize visual quality under the size constraint. At the inter-attribute level, we assign bit-widths to each attribute channel by formulating the combinatorial optimization as a 0-1 integer linear program, which can be efficiently solved. At the intra-attribute level, we divide each attribute channel into blocks of vectors, quantizing each vector based on the optimal bit-width derived at the inter-attribute level. Dynamic programming determines block lengths. Using the size estimator and MPQ, we develop a calibrated algorithm to identify optimal hyperparameters in just 10 minutes, achieving a 1.69$\times$ efficiency increase with quality comparable to state-of-the-art methods.

Computer Vision and Pattern Recognition,Multimedia

What problem does this paper attempt to address?

### What problems does this paper attempt to solve?

This paper aims to solve two main challenges faced by 3D Gaussian Splatting (3DGS) during compression:

1. **How to generate appropriate hyper - parameters according to a given size budget**: This requires establishing an accurate mapping relationship between file size and hyper - parameters.
2. **How to effectively select a set of hyper - parameters that can maximize visual quality under the premise of meeting the size budget**: This step requires quickly estimating the highest quality that each set of hyper - parameters can achieve.

Specifically, although existing compression methods can improve the compression quality to a certain extent, when adapting to different storage and transmission conditions, they are often unable to meet specific size limitations while maintaining high quality. Therefore, the author proposes a new framework named SizeGS to achieve the following goals:

- **Introduce a size estimator**: Used to predict the compressed file size, so as to efficiently search for hyper - parameters that meet the size budget.
- **Propose a hierarchical mixed - precision quantization scheme**: Optimize the visual quality under size constraints through a two - layer structure - between - attributes and within - attribute.
- **Design a calibration algorithm**: Find the hyper - parameter settings that meet the size budget and have the best quality within 10 minutes, with a performance 1.69 times higher than existing methods.

### Method overview

To achieve the above goals, the author adopts the following technical means:

1. **Size estimator**: Based on the offline compression framework MesonGS, define the size as a multivariate function of hyper - parameters (such as octree depth, retention rate, number of blocks, etc.), which simplifies the solution process.
2. **Hierarchical mixed - precision quantization (H - MPQ)**:
- **Between - attributes level**: Model the bit - width selection as a 0 - 1 integer linear programming problem (ILP) and solve it using the open - source PULP library.
- **Within - attribute level**: Divide each attribute channel into multiple blocks and use dynamic programming to determine the optimal length of each block to minimize information loss.
3. **Calibration algorithm**: Combine the size estimator and H - MPQ to quickly find the optimal hyper - parameter configuration through binary search and fine - tuning.

Through these methods, SizeGS can not only efficiently compress 3D Gaussian point clouds under a given size budget, but also significantly improve the visual quality after compression, which is suitable for application scenarios that need to adapt to different network bandwidths.