A short quiz on feature detection

Adrian F. Clark

1. How are BRIEF features matched?

   • exclusive OR

   • angle between vectors
   • OR
   • Euclidean distance

   BRIEF is intended to run in real time and uses strings of binary digits as the descriptor. The easiest way to measure dissimilarity of two bit-strings is to use exclusive-OR and then count the number of set bits.

2. What is the purpose of the Hough transform?

   • To find the parameters of lines, etc

   • To find edges in images
   • To calculate an accumulator array
   • To join together edges in images

   The purpose of the Hough transform is to estimate the parameters that describe some image feature -- though we have looked only at straight lines in this module.

3. Which of the following schemes is unsuitable for matching arbitrary features in images?

   • cross-correlation
   • matched filtering

   • the Laplacean

   • image differencing

   The Laplacean emphasizes isolated points in images, and is not suited to matching any other feature in an image.

4. Which of the following is NOT suitable for detecting corners?

   • FAST
   • Harris and Stephens

   • SIFT

   • Moravec

   SIFT tends to find features well away from corners, while the other detectors listed here were all devised to be corner detectors.

5. Which of the following transformations is not handled by SIFT?

   • contrast changes
   • rotations

   • reflections

   • scale changes

   SIFT is designed to handle moderate changes in scale and orientation, and is reasonably independent of contrast changes. However, it does not handle reflections.

6. What is the basis of the HOG feature detector?

   • a histogram of pixel values
   • a histogram of edge gradients

   • a histogram of edge orientations

   • a histogram of colour differences

   HOG stands for "histogram of oriented gradients", so the histogram is of edge orientations, also known as edge directions.

7. Why is the parametric Hough transform preferred to the non-parametric one?

   • It avoids problems then sin and cos are zero
   • It is more accurate

   • It avoids problems when the gradient becomes infinite

   • It is faster to calculate

   The biggest problem with the non-parametric Hough transform is that a vertical line has an infinite gradient.

8. How are peaks chosen from the Hough accumulator array?

   • Nearest to origin first
   • Smallest first
   • Smallest radius first

   • Largest first

   The peak with the most lines in the accumulator contributing to it is the most significant line, and this is the pixel with the largest value.

9. Which of the following can a Hough transform not be used to identify?

   • Straight lines
   • Circles

   • Faces

   • Ellipses

   Alas, the Hough transform cannot be used to identify faces because a face does not have a (small enough) set of parameters that describe it in an image.

10. What does one do with the Hough accumulator to find lines?

    • Each corner in the accumulator is a line in the image
    • Each 3 x 3 region in the accumulator is a line in the image

    • Each peak in the accumulator is a line in the image

    • Each line in the accumulator is a line in the image

    As the code in the notes illustrates clearly at the end of Chapter 5, peaks in the Hough accumulator correspond to lines in the image. This is because more image pixels have `voted' for that set of parameters than any other set.