Distinctive image features from scale-invariant keypoints

David G. Lowe (2004)

Key points

Invariant to scaling and location, partially invariant to illumination and viewpoint
Local: robust against occlusion, noise
Highly distinctive
Steps:
1. Scale-space extrema detection
2. Keypoint localization (reject poor contrast regions and edges, keep corners!)
3. Orientation assignment
4. Keypoint descriptor
Matching: approximate nearest-neighbor, compare to 2nd-nearest neighbor: 1st-nearest neigbor must be considerably closer for a reliable match