Lesson 09 · Book chapter pointer
Feature extraction
Read the chapter source: features.tex — do not treat this note as the book.
Learning goals
- Define a feature as a repeatable, distinctive, compact patch — not as “an interesting bit.”
- Explain detection vs description vs matching as three jobs.
- Say why corners beat flat walls for correspondence.
- List what invariance you actually have (scale, rotation, illumination) and what you wish you had.
- Connect matching quality to pose estimation and mapping later.
Teaching note
A feature is a contract: this region should be findable again, from a nearby view, and it should not look like half the rest of the image. Detection proposes where. Description writes a vector. Matching guesses which vectors are the same physical place. Students mash the three together and then wonder why “ORB is broken.” Name the stage that failed.
Corners and blobs exist because flat intensity is ambiguous. A Harris-style corner is a place where intensity changes in two directions; you can lock a patch there. An edge is only half a lock — it can slide. A blank wall is no lock at all. That is why warehouses with white paneling make visual odometry cry, and why you add tape or rely on lidar.
Descriptors try to be invariant to the nuisances you expect: a little rotation, a scale change, a brightness shift. Classic gradient histograms (SIFT-family thinking) and binary tests (BRIEF/ORB-family thinking) are two cultures. Neither is magic under viewpoint change that violates the assumption. Learned descriptors exist; they still fail on textureless plastic. Always look at the match image, not only at a score.
Matching is nearest-neighbor plus a refusal to be desperate. Ratio tests, mutual checks, and geometric verification (RANSAC on a homography or essential matrix) are how you throw away the confident wrong pair. One bad match can steer a pose more than ten good ones if you let it vote equally.
Classroom drill: take two frames of the lab, detect corners, match, and draw the outliers in a mean color. Then cover a textured poster and watch inliers vanish. That is the feature budget of your environment. Read the book chapter for the detectors and math this companion will not paste.
A match is a hypothesized correspondence \(p_i \leftrightarrow p_j\). Pose solvers consume inlier sets; they do not forgive a polluted one.