Skip to content
Free public beta until December 4, 2026: every course is open, no card needed. After that, the first chapter of each course stays free and full access is a one-time $39 ($19 for up to 100 founding members chosen for helpful feedback). Your progress is kept. How the beta works

Every way to recover depth and 3D structure

Stereo triangulation, monocular depth, and structure from motion -- three different routes to distance, each needing a different kind of input.

Ask first: how many images do you have, and do you know where the camera(s) were? Two calibrated cameras at a known baseline gives exact, metric depth by triangulation. A single image with no known geometry only gives relative depth, from a trained model's learned cues. Many uncalibrated images of the same static scene, taken by a camera moving through it, can recover both the 3D structure and every camera's own pose at once -- and, run live instead of as an offline batch, becomes SLAM.

Two (or more) cameras, known geometry

Term

disparity

See section →
The horizontal pixel-position difference between a matched point's location in a rectified stereo pair.

Example

A point 2 meters from the rig shows much larger disparity than one 50 meters away.

The raw per-point measurement a calibrated, rectified stereo pair gives directly.

See section →
Recovering a 3D point's location by intersecting the two camera rays that produced its matched image points.

Converts a matched correspondence plus known camera poses into an exact 3D coordinate.

See section →
A set of 3D points, typically with no connectivity between them, representing a scene's geometry.

What triangulating many correspondences at once produces.

One image, no known geometry

See section →
Estimating per-pixel depth from a single image, with no second camera view to triangulate from.

The only option with a single image and no calibrated baseline -- relative depth only, from learned cues, never an exact metric distance.

Many images, unknown camera positions

See section →
Recovering both a scene's 3D structure and every camera's pose from many overlapping images with no camera positions known in advance.

Recovers 3D structure AND every camera's pose at once, offline, from an unordered image set.

See section →
Tracking a moving camera's own pose while building a map of the scene, from a live stream, at the same time.

The same problem run live, frame by frame, with loop closure to correct drift instead of a one-shot batch solve.

Graph settings

Calculator