3D Gaussian Splatting

Peter Keum, GISP - King County Wastewater Treatment

What is it? & How do I use it?

June 17, 2025

This is 3D Gaussian Splatting model

Here is another..

And Lastly, 

CUGOS

King County Wastewater Treatment by Numbers

  • 2 million residents

  • 424 square miles

  • 383 miles sewer pipeline (2.5 inch to 14 ft)

  • 74 facilities (PS & RS)

  • 5 Treatment Plants: 180 MGD ~ Max. 900 MGD

Seattle, WA --> Medford, OR

WTD GIS Team

  • 7 GIS Specialists

Supervisor

WTD GIS

Maintain Sewer GIS Data

  • FIRS (Facility Information Retrieval System) Data
    • Maintenance Hole
    • Pipes
    • Pump Stations
    • Regulator Stations
    • Treatment Plants
    • Outfalls
  • Flow Monitors
  • CP Stations
  • Connection Points
  • Site Plan Data
  • 3D Models
  • Drone Collected Data
  • .... many many more

WTD

GIS

DATA

3D Gaussian Splatting

WA URISA - May 17, 2017

 1. What is it & why use it?

Traditional Photogrammetry Model (SfM)

Same model of 3DGS version

What is Gaussian Splatting?

  • Concept idea proposed back in 1991 - not able to be developed due to hardware limitation
  • SIGGRAPH - Special Interest Group on Computer Graphics and Interactive Techniques (computer graphics and interactive techniques) 

  3D Gaussian Splatting for Real-Time Radiance Field Rendering  (SIGGRAPH 2023)

"Rendering volumetric data directly without converting into a traditional surface model"

Photogrammetry - Measurement from a photo

SfM (Structure from Motion) techniques that can interpolate 3D models from overlapping images. 

Traditional Photogrammetry Model (SfM)

Credit: Olli Huttunen

Mesh based on TIN

Triangulated Irregular Network

What is Gaussian Splatting?

"Rendering volumetric data directly without converting into a traditional surface model"

What is Gaussian Splatting?

"It uses a mathematical function called the Gaussian, which makes this "continous visualization" in space possible" 

Credit: Olli Huttunen

"Spray Point" Approach

3D Gaussian

Splat

Credit: Olli Huttunen

Benefits of Gaussian Models

  • Does not use image texture
  • Can display reflections
  • Can show transparent surfaces
  • Good with thin structures
  • Photorealistic look
  • Great in low light
  • Runs in web browser
Credit: Olli Huttunen

Limitations of Gaussian Models

  • Covers only the angles that where captured
  • Accuracy
  • Zooming to details
  • No support on major 3D programs
  • Conversions to mesh formats
  • Lights and shadows are baked
  • Big file sizes
Credit: Olli Huttunen

Point Cloud

  • RGB pixel (DIM):
    • Dense Image Matching
  • Lidar - laser

3D Scanning

Mobile Phone

Drone

DSLR

360

Photogrammetry method

Credit: Olli Huttunen

 2. How do I make it?

How Gaussian Splatting Models Can be Generated?

Web services and mobile apps

Local training on your own Compter

SCANIVERSE

KIRI Engine

3Dpresso

Credit: Olli Huttunen

Example

Workflow

 #1 Video -> #2 Polycam (export ply) 

#3 SuperSplat (browser editor)

 #4 Upload back to Polycam and share

 3. How do I use it?

Integration of Lidar with Gaussian Splats

(SLAM) Simultaneous Localization and Mapping

Like drawing a map of a room AND figuring out exactly where you are on it, all at the same time. 

Example

of Gaussian Splat

SLAM Lidar

point cloud behind

Proof Of Concept

Lidar SLAM Handheld Unit

Video of Scan Path

Proof Of Concept

Lidar SLAM Handheld Unit

Examples

Proof Of Concept

Lidar SLAM Handheld Unit

Examples

Proof Of Concept

Lidar SLAM Handheld Unit

Examples

WHOA!

peter.keum@kingcounty.gov