A short quiz on feature detection

Adrian F. Clark

1. If a histogram has two peaks, where is the best place to put a threshold to separate foreground from background?

   • at the higher peak
   • half-way down the lower peak
   • at the lower peak
   • half-way up the higher peak

   • in the bottom of the dip between the two peaks

   In general, the best place to put a threshold is at the bottom of the dip between the two peaks -- this is essentially what Otsu's method does.

2. What was the major advance in the Harris and Stephens corner detector compared to Moravec?

   • it works with shifted regions around the corner

   • it considers the directions of the corner's derivatives

   • it works more quickly
   • it avoids detecting more than one corner at a single image location

   Harris and Stephens used calculus to determine the derivates at a corner, and from these it is able to determine their directions

3. Why is a recursive region labelling algorithm poor in practice?

   • it is difficult to program
   • it tends to overflow the computer's heap
   • it is too slow

   • it tends to overflow the computer's stack

   Recursive implementations of any algorithm save state on the program's stack. Recursive region-labelling algorithms make one recursive call for each pixel in a region, so if a region contains many pixels, stack overflow is likely.

4. What is characteristic that the Moravec corner detector uses to identify corners?

   • it finds minima in the maximum gradient of a region

   • it finds maxima in the minimum gradient of a region

   • it finds saddle points in the gradient of a region
   • it finds regions where the gradients in all directions are almost the same

   The Moravec corner detector calculates the gradient in four directions and retains the minimum of these. It then looks for local maxima in these minimum values to identify corners.

5. The Canny edge detector represents more or less the state of the art in detecting edges in images. Why is its output still of limited use?

   • it doesn't find all edges
   • the edges are still too noisy to be useful
   • it is too slow

   • edges are of limited value in practice

   In most cases, identifying the edges of objects is useful for segmenting them from the background; but edges are not much help in (say) identifying the object. For that task, it is much better to use the corners of features in the image.

6. What does Otsu's method do?

   • maximizes the within-class variance
   • minimizes the background variance

   • minimizes the within-class variance

   • maximizes the background variance

   Otsu's method minimizes the within-class variance as that is the best way of minimising incorrect background and object pixels; Otsu showed it is the same as maximizing the between-class variance.

7. What information does the Canny edge detector find about each piece of edge?

   • the direction
   • the magnitude
   • the magnitude, direction and width

   • the magnitude and direction

   Canny finds the magnitude and direction of each line by forming differences along each row and column.

8. Why does the last stage of the Canny edge detector employ two thresholds?

   • not all edges produce a response above both thresholds

   • not all edges produce a response below both thresholds
   • to reduce the amount of noise in an image
   • not all edges produce a response between both thresholds

   After the earlier stages, not all edge segments have a magnitude great enough to be above the upper threshold, though they should still be above the lower threshold.

9. Why does the last stage of the Canny edge detector perform hysteresis thresholding?

   • to reduce the effects of noise
   • to make edge segments only one pixel wide

   • to connect together edge segments

   • to make the output image binary

   The earlier stages of the Canny edge detector identify edge segments but they do not necessarily form a continuous edge. The purpose of the last stage is to connect together edge segments where there is evidence that they should be connected.

10. The first stage of the Canny edge detector involves smoothing with what?

    • a Laplacean mask

    • a Gaussian

    • a 3 x 3 blur mask
    • a Sobel mask

    The first stage of the Canny edge detector is to smooth the image (in the hope of reducing the number of false edges) with a Gaussian mask.