The encyclopedia · Software & IT · Technical decision · 1999–2004
SIFT made computer vision match images by local features that survive scaling and rotation
David Lowe's 1999 algorithm describes an image by stable keypoints instead of comparing whole pictures, so matching works across size, angle and clutter.
University of British Columbia
the move
In the late 1990s, matching an object across different photos was brittle: the same scene could be larger, rotated, dimmer, or partly hidden, and whole-image comparisons failed on any of those. David Lowe proposed extracting distinctive keypoints from each image and matching on those instead.
The keypoints are chosen at scale-space extrema, then filtered for stability, given a dominant orientation, and described by local gradient histograms. Because each feature describes only its own neighborhood, changes elsewhere in the image do not matter—the same object can be recognized even when the photo is taken from a different angle, distance or lighting.
The 2004 paper showed the approach could reliably identify objects in a database of many thousands of images, while handling partial occlusion and changing viewpoint. That combination—distinctiveness plus invariance—turned SIFT into the workhorse of early 2000s computer vision.
why it works
- Local features are robust to occlusion and clutter because the match only needs a handful of consistent keypoints.
- Scale-space keypoints and orientation assignment make matching independent of image size and rotation.
- High-dimensional descriptors are distinctive, so wrong matches are rare even in large databases.
- The method works on ordinary photographs without special hardware, which made it deployable everywhere.
what transfers
When exact comparison is impossible, find the smallest invariant unit you can match—robustness comes from local, redundant evidence, not from one global summary.
what came after
SIFT became one of the most cited papers in computer vision and a foundation for image stitching, panorama building, robot mapping, 3D reconstruction and early visual search. Its ideas fed into later feature methods (SURF, ORB) and deep-learning vision, and its patents shaped how vision features were licensed in the 2000s.
references
- Object Recognition from Local Scale-Invariant Features (ICCV 1999)
- Distinctive Image Features from Scale-Invariant Keypoints
spotted an error? The archive wants to know.