Feature Points

Corners, blobs, and other points in an image that can be reliably found again, together with a vector describing their surroundings.

Feature points are points in an image with distinctive texture that can be found again reliably even under a change of viewpoint or lighting — corners and blobs are typical examples. A feature point has two parts: a location (sometimes with a scale and orientation too), and a descriptor — a vector summarizing the appearance of the small patch around it, used to compare against other images. SIFT was proposed by David Lowe in 1999, with the full paper published in 2004; it’s robust to scaling, rotation, and lighting changes, with a 128-dimensional floating-point descriptor, and its patent expired in 2020. ORB was proposed by Rublee and colleagues in 2011, combining FAST corner detection with an improved binary BRIEF descriptor, making it much faster than SIFT and well suited to real-time systems. Feature points underlie visual SLAM, structure from motion, and image stitching, and the deep-learning era has also produced learned feature points like SuperPoint.

ExampleORB-SLAM (IEEE T-RO 2015) uses the same set of ORB features across all four of its stages: tracking, mapping, relocalization, and loop closure.

Also called

SIFT, ORB, Local Features, Keypoints

Related

Feature Matching · Visual SLAM · ORB-SLAM3 · SuperPoint / SuperGlue / LightGlue · Keypoint Detection · Structure from Motion

Sources

Wikipedia: Scale-invariant feature transform
Wikipedia: Oriented FAST and rotated BRIEF
ORB-SLAM: a Versatile and Accurate Monocular SLAM System