Feature Selection

Feature selection is the stage between detection and description that culls the raw set of keypoints down to a smaller, better-distributed subset. Detectors like Harris or DoG often return many keypoints concentrated in high-texture regions (buildings, foliage) while leaving other regions empty. A well-distributed set of keypoints gives geometric estimators (homography, fundamental matrix) more leverage than a clustered set of the same size.

Intuition

Raw detection responses are noisy — thousands of keypoints cluster on a brick wall while the plain sky beside it has none. Downstream RANSAC cares not just about how many inliers exist but where they are: four inliers well-spread across the image are better than forty inliers packed into one corner. Feature selection is the practical step that enforces spatial diversity and keeps the keypoint count manageable for matching.

Notes

• 4.2.1. ANMS — Adaptive Non-Maximal Suppression: spatially balanced keypoint selection

See Also

• 4.1. Feature Detection — the raw responses that selection operates on
• 4.3. Feature Description — the next stage that consumes selected keypoints