Principle:Kornia Kornia Local Feature Detection
| Knowledge Sources | |
|---|---|
| Domains | Vision, Feature_Detection |
| Last Updated | 2026-02-09 15:00 GMT |
Overview
Technique of detecting repeatable interest points and computing local feature descriptors at multiple scales for image matching and recognition.
Description
Local feature detection identifies distinctive image locations (keypoints) that are stable across viewpoint and illumination changes. The classical pipeline involves: (1) building a scale-space pyramid, (2) computing a response function (Harris corners, Hessian blob detector, or DoG), (3) applying non-maximum suppression to select peaks, (4) estimating affine shape and dominant orientation, (5) extracting a local descriptor (SIFT, HardNet, SOSNet) from the normalized patch. Kornia's ScaleSpaceDetector implements this as a differentiable nn.Module pipeline.
Usage
Use when you need sparse, repeatable keypoints with descriptors for matching, tracking, or recognition. Preferred when interpretability and control over feature selection matter more than matching density.
Theoretical Basis
Scale-space representation:
L(x, y, sigma) = G(x, y, sigma) * I(x, y)
Keypoints are extrema of the Difference of Gaussians (DoG):
D(x, y, sigma) = L(x, y, k*sigma) - L(x, y, sigma)
Non-maximum suppression selects local maxima in 3x3x3 neighborhoods. Keypoint localization refines positions by fitting a quadratic to the DoG response.
Output: Local Affine Frames (LAFs) encoding position, scale, and orientation.