3 《3D Gaussian Splatting: From Theory to Real-Time Implementation》第三级：压缩、轻量化与存储优化 （一）

目录

第一部分：原理详解

1.1 Scaffold-GS原理：神经高斯与锚点的空间层次结构

1.1.1 神经高斯与锚点的空间层次结构

1.1.2 局部感知神经解码与视锥剔除机制

1.1.3 锚点层级扩展与多尺度场景覆盖

1.2 可微分量化：Laplacian-based Rate Proxy与熵约束优化

1.2.1 Laplacian率失真代理模型

1.2.2 熵约束端到端优化框架

1.2.3 非均匀量化与自适应码本优化

1.3 4D锚点网格的时间扩展：动态场景中的时空梯度累积策略

1.3.1 时空可变形锚点网格

1.3.2 时空梯度累积与关键帧选择

1.3.3 时间一致性正则化与流场监督

2.1 Voxel-GS的游程编码：Octree + Run-length Coding的混合方案

2.1.1 八叉树空间划分与占据编码

2.1.2 游程编码压缩连续空区域

2.1.3 GPU友好解码与流式加载

2.2 量化感知训练(QAT)：对协方差、SH系数、不透明度的混合精度量化

2.2.1 几何参数（协方差）的约束感知量化

2.2.2 外观参数（SH系数）的知觉量化

2.2.3 不透明度与辅助属性的非对称量化

2.3 紧凑表征的神经替代：Compact3DGS的哈希网格编码与MLP预测

2.3.1 多分辨率哈希网格编码

2.3.2 轻量级MLP属性预测

2.3.3 渐进式哈希网格剪枝与激活稀疏化

3.1 渐进式加载：基于LOD(Level-of-Detail)的高斯点云流式传输

3.1.1 屏幕空间误差驱动的LOD选择

3.1.2 八叉树LOD层次构建与节点编码

3.1.3 流式传输协议与预取策略

3.2 针对边缘设备的剪枝策略：基于重要性采样的高斯子集选择

3.2.1 高斯重要性度量与显著性评估

3.2.2 重要性采样与确定性剪枝算法

3.2.3 运行时动态剪枝与质量-能耗自适应

第二部分 伪代码

Algorithm 1: Anchor-based Gaussian Generation and Neural Decoding

Algorithm 2: Frustum-aware Visibility Culling and Lazy Instantiation

Algorithm 3: Octree Hierarchy Construction for Multi-scale Coverage

Algorithm 4: Laplacian-based Rate-Distortion Optimization

Algorithm 5: Entropy-Constrained Quantization-Aware Training

Algorithm 6: Adaptive Non-uniform Codebook Optimization

Algorithm 7: Spatio-temporal Deformable Grid Construction

Algorithm 8: Temporal Gradient Accumulation with Key Frame Selection

Algorithm 9: Temporal Consistency Regularization and Flow Supervision

Algorithm 10: Sparse Octree Construction and Occupancy Encoding

Algorithm 11: Run-Length Encoding for Empty Node Compression

Algorithm 12: GPU-friendly Octree Decoding and Coalesced Access

Algorithm 13: Geometry-aware Covariance Quantization

Algorithm 15: Asymmetric Opacity and Auxiliary Attribute Quantization

Algorithm 16: Multi-resolution Hash Grid Encoding

Algorithm 17: Lightweight MLP Attribute Prediction

Algorithm 18: Progressive Hash Grid Pruning and Sparsification

Algorithm 19: Screen Space Error LOD Selection

Algorithm 20: Octree LOD Hierarchy Construction with Delta Coding

Algorithm 21: Streaming Protocol with Prefetching and Delta Updates

Algorithm 22: Gaussian Importance Assessment for Edge Devices

Algorithm 23: Hierarchical Importance-based Gaussian Pruning

Algorithm 24: Runtime Adaptive Pruning with Quality-Energy Tradeoff

第一部分：原理详解

1.1 Scaffold-GS原理：神经高斯与锚点的空间层次结构