Distinctive Image Features from Scale-Invariant Keypoints

Inside the paper

The problem

Matching an object across images is difficult when its size, orientation, viewpoint, or lighting changes. A useful local feature must be repeatable enough to find again and distinctive enough to reject unrelated image regions.

The idea

SIFT identifies stable image locations at multiple scales and describes the local gradient patterns around them in a normalized coordinate frame. The resulting features can be compared across images and checked for geometric agreement.

How it works

• The method searches a scale-space built from differences of Gaussian-blurred images for candidate extrema in position and scale.
• It refines each candidate and removes unstable points, including weak responses and points poorly localized along edges.
• Local gradient directions determine a reference orientation, so the descriptor can be expressed relative to the feature rather than the image axes.
• The descriptor summarizes nearby gradient orientations in spatial bins, and matching uses nearest neighbors followed by geometric verification.

What it showed

The paper tests feature repeatability and matching under changes including scale, rotation, illumination, and viewpoint. It also demonstrates object recognition in clutter by grouping consistent matches and fitting an affine transformation.

Read it for

Start with Section 3 and Figure 1 for scale-space construction, then Section 6 for the descriptor and its tests. The nearest-neighbor indexing details in Section 7.2 can wait until you need to build a large database.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.