3DGS.zip: A survey on 3D Gaussian Splatting Compression Methods(2)
Abstract
An ongoing survey has been initiated to examine the current state of 3DGS compression techniques, with an emphasis on their statistical performance across diverse benchmarks. This effort seeks to enhance comparability by compiling key statistics of various compression methods in a tabular layout.
These datasets under evaluation encompass TanksAndTemples, MipNeRF360, DeepBlending, and SyntheticNeRF.
For each method, we present the PSNR value, SSIM index, LPIPS metric, and output file size in megabytes (MB), as reported by various authors.
Figure
Figure 8
gsplat
This approach incorporates 3DGS as a method of Markov Chain Monte Carlo (3DGS-MCMC) utilizing compression techniques derived from the 'Self-Organizing Gaussians' paper and achieving more efficient Gaussian splat creation.
This open-source library incorporates CUDA-accelerated differentiable rasterization of 3D gaussians with Python bindings.
The library draws on the SIGGRAPH paper titled "3D Gaussian Splatting for Real-Time Rendering of Radiance Fields", however, gsplat stands out as it is not only faster but also significantly more memory-efficient. Furthermore, it now boasts an expanding array of new features.
Figure 9
Compact-3D: Compression of Gaussian Spray Radiance Fields using Vector Quantization
This approach employs a quantization method based on the K-means algorithm to partition Gaussian distributions in 3DGS, as numerous Gaussians may share identical Gaussian distributions.
仅存储了一个小型代码本，并在每个高斯分布中同时包含了其索引信息，在学习到光照场的过程中带来了显著的减少，在渲染过程中对内存占用的影响也得到了优化。
Additionally, the indices are further reduced by sorting the Gaussians dependent on one of the quantized parameters and **employing a technique akin to Run-Length-Encoding (RLE) for storing the indices.
To minimize the count of Gaussians, this approach incorporates a regularizer to promote zero opacity settings, thus removing Gaussians whose opacity falls below a predefined cutoff level.
Figure 10
Compact 3D Scene Representation via Self-Organizing Gaussian Grids
Compressing 3D data presents significant challenges, yet a multitude of highly effective solutions have been developed for the compression of 2D data, including a wide range of visual media such as images.
The authors introduce a new method to structure 3DGS parameters into a 2D grid, which significantly reduces storage requirements while maintaining the visual quality.
This organization capitalizes on visual redundancy within natural visual scenes.
They present as an extremely efficient parallel sorting algorithm, known as PLAS, which organizes the Gaussian parameters across a two-dimensional grid, preserving the local neighborhood relationships while ensuring spatial coherence.
Such a solution stands out as cutting-edge, as it fails to handle a 2D grid with millions of points efficiently.
While training, a smoothness loss is utilized to maintain local smoothness within the 2D grid, which improves data compressibility.
The central realization is that smoothness must be maintained during training for achieving effective data compression.
Figure 11
Implicily defined Gaussian splatting utilizing a highly efficient multi-level tri-plane representation
This method presents a composite model for radiance fields in splatting-based rendering, employing Gaussian primitives that are divided into distinct point clouds and implicit feature attributes.
The attribute features are represented through a combination of a multi-resolution, multi-level tri-plane architecture and a residual-based rendering pipeline.
It adopts a hierarchical training framework for the collective optimization of point cloud data and plane features, generating a comprehensive feature set starting from basic geometric attributes and incorporating increasingly detailed information.
Space-based regularization and a bootstrap-based approach are subject to improving the consistency and stability of the Gaussian attributes during the training process.
Figure 12
The MesonGS framework introduces an efficient method for post-training compression of 3D Gaussian distributions through an optimized attribute transformation process.
MesonGS implements universal Gaussian pruning by assessing the significance of Gaussians via forward propagation, taking into account both view-dependent and view-independent features.
This converts rotation quaternions into Euler angles with the aim of minimizing memory usage, employing the region adaptive hierarchical transform (RAHT) to cut down on entropy within critical attributes.
通过块量化技术,Block quantization被应用于属性通道中,并将其划分为多个区域,然后对各个区域分别执行独立的量化处理,其中向量量化被用来压缩那些不重要的属性
The geometry is being compressed by an octree, and all elements are being packed by the LZ77 codec.
A finetune scheme is implemented post-training to restore quality.
Figure 13
Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis
The authors have developed an efficient 3D Gaussian splat representation, which is composed of three primary components.
基于敏感度的聚类方法中
The quantization-aware fine-tuning method restores lost information through the tuning of model parameters at reduced bit-rates, achieved via quantization-aware retraining.
3. entropy-based encoding, which makes use of spatial redundancy through entropy and run-length encoding techniques by linearizing 3D Gaussians along a space-filling curve.
Furthermore, a renderer for the compressed scenes utilizing GPU-based sorting and rasterization is proposed, enabling real-time novel view synthesis onlow-end devices.
Figure 14
F-3DGR: Decomposed Coordinate Systems and Representation Models for Surface Reconstruction
The paper presents an innovative 3D Gaussian compression method by employing structured coordinates and decomposed representations via factorization techniques.
This method is inspired by the techniques of tensor or matrix factorization. It generates 3D coordinates through the tensor product of one-dimensional or two-dimensional coordinate systems to construct a three-dimensional coordinate system. The aim is to improve spatial efficiency.
This method expands factorization by incorporating attributes such as color, scale, rotation, and opacity, effectively reducing the model size without losing key features.
Binary masks are typically utilized to remove ineffective Gaussians, thereby significantly accelerating both training and rendering processes的速度提升.