Feature Detection

Feature detection finds image locations — keypoints or interest points — that are distinctive enough to be re-identified across images of the same scene. A good detector is repeatable (the same physical point yields a keypoint in every view), localized (precise to a pixel or better), sparse (few enough points to be matched efficiently), and invariant to nuisance transformations (rotation, scale, illumination, moderate viewpoint change).

The standard pipeline pairs a detector with a descriptor and a matcher to produce the correspondences that feed homography estimation, epipolar geometry, and structure from motion.

Intuition

A pixel on a plain wall looks like every other pixel on the plain wall — matching it across images is hopeless. A pixel at the corner of a window, by contrast, has a unique local neighborhood that cannot be confused with its surroundings. Detection is the problem of locating these distinctive pixels automatically, before any descriptor is computed.

The design tension is between distinctiveness (pick points with unique neighborhoods) and invariance (pick points whose response is stable under zoom, rotation, and illumination change). Early detectors (Harris) optimize the first; modern detectors (DoG, Hessian-Laplace, SuperPoint) add scale selection and learned robustness on top.

Notes

• 4.1.1. Harris Corner Detection — the classical rotation-invariant corner detector based on the second-moment matrix
• 4.1.2. Scale-Invariant Detection — Difference-of-Gaussians extrema for scale-invariant keypoints (the detector stage of SIFT)

See Also

• 4.3. Feature Description — the descriptor stage that turns detected keypoints into matchable vectors
• 4.4. Feature Matching — the correspondence stage that pairs descriptors across images