Tutorials

Accepted Tutorials

The following tutorials will be presented at MMM 2027.

Optimizing 3D Scene Representations: From Gaussian Splatting Compression to Self-Organizing Grids

Presenters

Abstract

While 3D Gaussian Splatting (3DGS) has revolutionized real-time neural rendering, its explicit representation introduces significant storage and memory bottlenecks. This tutorial provides a comprehensive guide to state-of-the-art techniques for efficiency, storage optimization, and structured data organization in 3DGS. We analyze the foundational differences between compaction (primitive reduction) and compression (quantization and entropy coding), reviewing recent frameworks such as HAC, Compact3D, and Scaffold-GS. Bridging computer graphics and data visualization, we introduce advanced grid-based sorting methods—specifically Fast Linear Assignment Sorting (FLAS)—to show how arranging unstructured 3D attributes into sorted 2D grids exploits inherent data redundancies. Finally, we synthesize these concepts into the "Self-Organizing Gaussians" (SOG) workflow, demonstrating how standard image codecs can be leveraged for highly efficient 3D scene representation without sacrificing rendering fidelity.

Contents

Part 1: Foundations of Efficient 3DGS (approx. 60 mins)
  • 3DGS Fundamentals: Overview of explicit scene modeling via 3D Gaussians, covariance matrices, and rasterization.
  • 3DGS Extensions: Overview of current variants and novel representations.
  • The Storage Bottleneck: Analyzing the high memory costs of explicit representations.
  • Compaction Strategies: Techniques for removing non-contributing splats and merging redundant geometry.
  • Attribute Compression: Implementing vector quantization and entropy coding to minimize precision footprints.
Part 2: Advanced Sorting and Grid Structures (approx. 60 mins)
  • Grid-based Sorting: Introduction to arranging high-dimensional data onto structured 2D grids.
  • Linear Assignment Sorting (LAS): Data arrangement as a cost-minimization problem.
  • Fast LAS (FLAS): Optimizing sorting speed for massive datasets using block-based solvers.
  • Permutation Learning: Creating sorted grid layouts with gradient-based optimization.
Part 3: Self-Organizing Gaussians & Integration (approx. 60 mins)
  • Self-Organizing Gaussians (SOG): Mapping 3D Gaussian attributes onto sorted 2D image grids.
  • Unified Pipeline: Leveraging sorting to smooth data signals, enabling highly efficient standard image compression.
  • Q&A