AI and machine learning
Stereo vision lab
Place two virtual cameras in a 3D scene, change baseline and focal […]
3D reconstruction from images
From a handful of photos to a point cloud and a navigable 3D scene, with the camera poses shown. The same geometry as the stereo lab, extended to many views.
A small version
A small version of the project. The full one is linked below.
How it works
- 01Build a small shape from towers, then walk around it and take photos from wherever you like. The algorithm never sees your shape, only the photos.
- 02Each photo is reduced to a black-and-white outline: the object is somewhere inside it. The shape is rebuilt by carving. A block of space survives only if it falls inside the outline in every photo that sees it, so each new viewpoint cuts away more of what is not there. With one photo the result is a long prism along the viewing direction, because depth is unknown. With a few well-spread photos it closes in on the real shape.
- 03Red blocks are material no photo has ruled out yet. They show what a missing viewpoint costs, such as the hollow under an overhang that no outline can reveal. Real systems also match points between photos to recover the camera positions; here the positions are known, which keeps the idea in focus.
What it will include
- 01Sparse point cloud and camera poses computed offline, viewable and orbitable in the browser.
- 02Click a camera to see the photo it took.
- 03A Gaussian splat of a real scene as an upgrade.
- 04A write-up linking it to the stereo lab: two views, then many.
More in this category
AI and machine learning
Sound source localization
Find where a sound comes from using the delay between microphones […]
AI and machine learning
HARNAIS SSI
A cybersecurity harness for AI coding agents such as Claude Code […]