24AI2002 - Corner Detection Quiz

Identify the correct description of a "corner" in an image.

• A point where image intensity is constant

• A point where two edges meet with intensity changing in one direction only

• A point where intensity changes in two or more directions, such as the intersection of edges

• A point with no gradient in any direction

What does it indicate when both x and y gradients around a point are large?

• A flat region

• An edge

• A corner

• Noise

What role do the eigenvalues λ₁ and λ₂ of the Harris matrix (M) play in corner detection?

• Both small → Edge; Both large → Flat

• One large and one small → Corner

• Both large → Corner; One large, one small → Edge; Both small → Flat

• Eigenvalues are not used

The Harris corner response function is given by R = det(M) – k·trace(M)². What kind of regions result when R is high?

• Flat regions

• Edges

• Corners

• Regardless of R, it's never a corner

What is the purpose of non-maximum suppression in corner detection?

• To blur the image

• To keep only the strongest corner in a local neighborhood

• To convert edges to corners

• To equalize the histogram of corner values

In the Harris corner detection algorithm, what does a high corner response map indicate?

• Likely corners

• Likely edges only

• Flat areas

• Regions of high noise

What is the main difference between edges and corners in terms of gradient change?

• Edges have a change in intensity in multiple directions, while corners have change in one direction

• Corners have intensity change in two or more directions; edges only in one

• Edges are detected by color; corners are detected by shape

• Corners are always brighter than edges

Which parameter controls the sensitivity of the Harris corner response (the k-value)?

• The kernel size

• The threshold value

• The scalar constant 'k' in R = det – k·trace²

• The number of detected corners

How is the matrix M derived in Harris corner detection?

• From the Fourier transform of the image

• From second derivatives along each pixel coordinate

• From intensity gradients in x and y, accumulated in a local window

• From RGB color space segmentation

Design an improvement: how would you adapt the Harris detector to better handle scale changes (size of features)?

• Use only one threshold for all scales

• Apply the detection on a single resolution image

• Use multi-scale image pyramids and apply the detector at multiple scales

• Ignore gradient information