. The structure tensor M built from image gradients is exactly analogous to which earlier concept?
Both are 2x2 covariance-style matrices; Lesson 6 built one from pixel coordinates, this lesson builds the same kind of matrix from local image gradients.
. What do both eigenvalues of the structure tensor being large indicate about a point?
A corner has strong gradient in every direction, which is exactly what two large eigenvalues of M mean.
. Why are Harris and Shi-Tomasi corner detectors NOT scale-invariant, motivating SIFT?
A sharp corner becomes a gentle curve at a fixed window size when the image is zoomed out enough, so a fixed-scale detector misses it.
. What does Lowe's ratio test do when matching SIFT descriptors?
Comparing the best match's distance to the second-best's rejects ambiguous matches where two candidates are nearly equally good.