How sure must a match be?
Each point of the first photo is paired with its nearest descriptor in the second. A pair is kept when that nearest one is clearly nearer than the second nearest. Green pairs land where an independent reference says they should, red ones do not. Tap a point to read its two distances.
Try this
Move the threshold from 0.75 up to 1.0 and watch the red lines pour in, then down to 0.5 and count the correct ones you lose.
Where this lab comes from
Lowe's Ratio Test: Matching Features You Can TrustHow SIFT matches points between two photos: pair each descriptor with its nearest neighbour, then keep the pair only when the second-nearest is clearly farther. On the Barcelona harbour pair, scored against a homography fitted by a different detector, the ratio test at 0.75 keeps 440 matches of which 420 are right; an absolute distance threshold tuned to keep the same 440 gets 391 right. Built in NumPy and checked against OpenCV's BFMatcher.
More in edges and features
- An edge, and its two derivatives — One row of pixels, with the first and second derivative stacked underneath.
- Canny, one stage per panel — Smoothing and the two hysteresis thresholds, on a real photo.
- Flat, edge, or corner — Drag a window over a photo; two eigenvalues deliver the verdict.