What is Gaussian Splatting?
Module Overview
| Est. time | 30 minutes |
| Type | Foundations |
| Prerequisites | 3D: Camera and Perspective |
| Read first | Gaussian Splatting technique page |
Introduction
Imagine you have a video of a crime scene — a drone flying over a building, a bodycam walking through a room, or a smartphone panning across a street. You need to turn that footage into a 3D space you can navigate, measure in, and present to others.
Traditional photogrammetry would reconstruct this as a mesh — a surface made of triangles. It's clean, measurable, and familiar. But meshes struggle with fine details like foliage, reflective surfaces, and complex lighting. They require significant processing time and cleanup.
Gaussian Splatting takes a different approach. Instead of building a surface, it represents the scene as a cloud of thousands of tiny, overlapping 3D blobs (Gaussians), each with its own position, size, shape, color, and opacity. When rendered, these blobs blend together to create a photorealistic view from any angle.
Think of it like a pointillist painting — close up you see individual dots, but step back and they form a coherent, detailed image. The difference is that these dots exist in 3D space and can be viewed from any perspective.
A Brief History
Gaussian Splatting is a recent technique, but its roots go back decades:
| Year | Development | Significance |
|---|---|---|
| 2023 | 3D Gaussian Splatting paper (Kerbl et al.) | Introduced the technique — real-time radiance field rendering from multi-view images |
| 2022 | Instant NGP (Müller et al.) | Multi-resolution hash encoding made NeRF training fast enough for practical use |
| 2020 | NeRF (Mildenhall et al.) | Introduced neural radiance fields — the idea of representing a scene as a continuous 5D function |
| 2016+ | COLMAP (Schönberger et al.) | Robust Structure-from-Motion pipeline — the standard way to get camera poses for both NeRF and 3DGS |
| 2000s | Point-based rendering | Earlier work on rendering point clouds with surface splatting — the conceptual ancestor |
| Classic | Structure from Motion / Photogrammetry | The foundational idea: reconstruct 3D from overlapping 2D images |
The key insight of the 2023 paper was real-time optimization and rendering. Previous radiance field methods (NeRF) required seconds to render a single frame — 3DGS achieved real-time (30+ fps) while maintaining or exceeding visual quality.
Why This Matters for Visual Investigation
The Opportunity
Gaussian Splatting offers several advantages for investigation workflows:
- Speed — A splat can be trained in minutes to hours, not days. This means faster turnaround for scene visualization.
- Visual fidelity — Fine details (grass, trees, fabric, water) that trip up mesh reconstruction are often handled naturally.
- View synthesis — You can generate novel views that were never captured, useful for verifying witness perspectives.
- Camera path validation — The splat's camera alignment can cross-reference your own camera solve.
The Caveat
Gaussian Splatting is a visualization tool, not a measurement tool by default.
The technique optimizes for rendering quality, not geometric accuracy. Splats can:
- Hallucinate detail in areas with sparse coverage
- Drift in scale if not properly constrained
- Produce artifacts on reflective or transparent surfaces
- Lack the clean topology needed for simulation or 3D printing
This means Gaussian Splatting is best used as context — a way to understand a scene spatially — combined with traditional mesh techniques for measurement-oriented tasks.
Where It Fits in the Workflow
In a typical visual investigation, Gaussian Splatting sits alongside photogrammetry:
Raw footage
↓
COLMAP / Metashape ───→ Camera solve + sparse point cloud
↓ ↓
Mesh reconstruction Gaussian Splatting
↓ ↓
Clean geometry Photorealistic visualization
↓ ↓
Measurement & analysis Context & communication
↓ ↓
→ Combined in Blender ←
Key Concepts at a Glance
Before diving into the pipeline, get familiar with these terms:
| Term | What it means |
|---|---|
| Gaussian | A 3D blob defined by position (mean), shape (covariance), color (spherical harmonics), and opacity |
| Splatting | The rendering technique — projecting each Gaussian onto the image plane and compositing them |
| Radiance Field | A representation that encodes both color and light direction at every point in space |
| Anisotropic | Not uniform in all directions — each Gaussian can be stretched/rotated independently |
| SfM / COLMAP | Structure from Motion — the camera solving step that provides initial poses for training |
| View Synthesis | Generating new images from viewpoints not present in the training data |
Knowledge Check
- What is the fundamental difference between Gaussian Splatting and mesh-based photogrammetry?
- Why does Gaussian Splatting handle fine details (like foliage) better than traditional meshes?
- What is the main limitation of Gaussian Splatting for forensic measurement?
Answers in the Resources section.
Summary
- Gaussian Splatting represents scenes as clouds of 3D Gaussians, not meshes
- It was introduced in 2023 and builds on earlier radiance field (NeRF) research
- It offers fast training, real-time rendering, and high visual fidelity
- For visual investigation, it's best used for spatial context alongside mesh-based measurement
- The pipeline involves: capture → SfM alignment → training → cleanup → integration
Resources
Knowledge Base Pages
- Gaussian Splatting technique page — Quick reference with methodology docs
- Photogrammetry technique page — The mesh-based counterpart
- 3D: Camera and Perspective — Foundational concepts
Key Papers
- Kerbl et al. (2023) — 3D Gaussian Splatting for Real-Time Radiance Field Rendering
- Mildenhall et al. (2020) — NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Videos
- Gaussian Splatting: End-to-End Workflow — linked in the KB technique page
- Capture and Reconstruction for 3DGS — linked in the KB technique page
For Reference
Knowledge Check Answers
- GS stores the scene as a point-based radiance field (soft Gaussians), while photogrammetry builds a triangle mesh surface
- GS doesn't require clean topology — every pixel is a blend of Gaussians, so complex structures like foliage are represented naturally without explicit geometry
- GS optimizes for visual quality, not geometric accuracy — splats can hallucinate detail, drift in scale, and lack the precision needed for direct forensic measurement without cross-referencing