🧭 Corners, Blobs and Keypoints

About this simulation

Before two photos of the same scene can be aligned, stitched into a panorama, or used to triangulate a 3D point, an algorithm has to find distinctive, repeatable keypoints in each image and describe them well enough to recognise the same physical point again from a different angle. This lab shows a synthetic reference photo on the left and a second "photo" of the same scene on the right — rotated, rescaled and made noisy, just like a real camera moving and re-shooting the same subject.

How it works

• Harris corners (red) look for points where the image gradient changes sharply in two directions at once — the corner of a window frame, not an edge or a flat wall. Harris only reports a location: it has no idea of scale or orientation, so it struggles once the image is zoomed or heavily rotated.
• SIFT blobs (green) search a scale-space of blurred copies of the image for blob-like regions that stay stable as size changes. Each keypoint carries a scale (the ring radius shown here) and a dominant orientation, which is what makes SIFT robust to zoom and rotation.
• ORB keypoints (blue) combine a fast corner detector (FAST) with an assigned orientation and a compact binary descriptor (BRIEF-style), giving most of SIFT's rotation robustness at a fraction of the computational cost — the reason ORB is popular in real-time and mobile vision.
• A keypoint's strength (its cornerness / blob response) is degraded by sensor noise; weak keypoints vanish first, which is exactly why noisy or low-contrast photos yield fewer reliable matches.

Controls

• Detector — switch between the three keypoint families and see which one survives the current viewpoint change.
• Viewpoint rotation / scale — simulates the camera rotating and moving closer or farther for the second photo.
• Sensor noise — degrades keypoint strength in the second image, causing weaker keypoints to drop out.
• Show matches — draws lines between keypoints that were successfully matched between the two images. Drag to rotate the 3D view, scroll to zoom.

Did you know?

SIFT (Lowe, 1999/2004) was patented for two decades and that patent expiring in 2020 is a big part of why it became a free default in OpenCV; ORB (Rublee et al., 2011) was explicitly designed at Willow Garage as a fast, unencumbered alternative for real-time robotics and mobile applications.