AI and machine learning
3D reconstruction from images
From a handful of photos to a point cloud and a navigable 3D scene […]
Stereo vision lab
Place two virtual cameras in a 3D scene, change baseline and focal length, and watch disparity and depth change. Then recover depth from the two images with classical matching and a small neural network.
A small version
A small version of the project. The full one is linked below.
How it works
- 01Two virtual cameras sit side by side, a baseline B apart, looking the same way. Each renders what it sees, so every object lands at a slightly different horizontal position in the two images. That shift is the disparity d, and depth follows from Z = f · B / d.
- 02The disparity map is found the way a classic stereo system finds it: for every pixel of the left image, slide a small window along the same row of the right image and keep the shift where the two windows look most alike (sum of absolute differences). No model is involved yet.
- 03The true depth is also rendered, so the right-hand picture shows exactly where matching is wrong. Remove the texture or add noise and flat regions fail; widen the baseline and near objects leave the search range. A learned network, trained on scenes from this renderer, is the next step.
What it will include
- 013D scene with two adjustable cameras, frustums, epipolar lines and live depth formula.
- 02Depth error against distance for the chosen setup.
- 03Block matching running on the GPU, compared with a learned disparity network.
- 04The network trained on scenes rendered by the same engine, with true depth as ground truth.
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 […]