ProjectsExperiments
Gaussian Splatting research for 3D web visualisation
Research into Gaussian Splatting to reconstruct 3D objects and scenes from photographs and display them interactively in the browser with Three.js.
- Status
- Finished
- Period
- 2026.08 — 2026.09
- Type
- Research
- Role
- Solo project, 3D development, Tooling and pipeline, Research
- Technologies
- Gaussian Splatting
- Three.js
- COLMAP
- gsplat
- Development log
- 1 entry
Context and goal
I researched how to create 3D representations of objects and scenes from photos taken with my phone. The goal was to understand how Gaussian Splats are made and to explore how to bring them into interactive web experiences with Three.js.
Technical description
I explored each stage of the process:
- Capture: photos of the object from many viewpoints.
- Camera reconstruction: with COLMAP and GLOMAP, which estimate the position and orientation of each photo and produce an initial point cloud.
- Training: with gsplat, which starts from that point cloud and optimises the Gaussians by comparing what they render against the original photos.
- Web viewing: in Three.js with the @mkkellogg/gaussian-splats-3d library, which sorts the Gaussians by depth and draws them in real time in the browser.
I captured the objects with my phone's camera on a makeshift set on my desk: the object in the centre and photos all around it, at different heights, to cover it from every angle.
Development log
First test with Gaussian Splatting and Three.js
A crocheted dinosaur, reconstructed with COLMAP and gsplat, with an inner light and a deformation that pulses with it.