A short quiz on low-level vision

Adrian F. Clark

1. In a real-time implementation of the Sobel operator, it is important to keep the number of multiplications as small as possible. What is the smallest number of multiplications that are required to convolve each image region with one of the Sobel masks?

   • 1
   • 6
   • 9

   • 0

   Multiplication of any number by zero is obviously zero, and multiplication by unity just yields the number. Multiplication by two can be done (for integer operands) by a bitwise left-shift, so convolution with the Sobel masks do not involve any multiplications at all.

2. In the broken biscuit identifier explored in lectures, what is its most serious problem?

   • it doesn't identify biscuits that aren't aligned with the image edges

   • non-uniform lighting

   • inadequate testing
   • it doesn't identify overlapping biscuits

   The most significant shortcoming is in thresholding the biscuits from the background because of non-uniform lighting, even though your lecturer went to some effort to make it as uniform as possible.

3. What does Otsu's method do?

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

   • minimizes the within-class 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.

4. Why is simple thresholding not especially effective at locating light features in an image?

   • thresholding is a local operation

   • changes in illumination change the lightness of features

   • thresholds are difficult to determine
   • thresholding is a global operation

   The main problem with feature detection using thresholding is that changes in illumination can cause (say) a white object to appear grey.

5. A system for identifying broken custard cream biscuits looks for rectangular regions that are aligned with the edges of images. Why might this be a bad idea?

   • whole biscuits might not be aligned with the edges of images

   • it's not really a bad idea
   • the shape of the biscuit doesn't matter, only its appearance
   • broken biscuits might not be aligned with the edges of images

   If the buscuit is not aligned with the edges, its axis-aligned bounding box will be much larger than an oriented bounding box and this might incorrectly identify the biscuit as being broken -- see figure 5.10 in the lecture notes.

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

   • it is too slow

   • it tends to overflow the computer's stack

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

   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.

7. An image of a room contains a computer with a display. What feature or features would help you detect the display?

   • rectangularity
   • corners

   • a combination of rectangularly and aspect ratio

   • SIFT features

   The best approach would be to combine rectangularity (to detect the rectangular feature) and aspect ratio, as computer monitors have a 4:3 or 16:9 aspect ratio. SIFT would be poor and corners do not carry enough information in this context.

8. A shape descriptor consists of the distance from the middle of a feature in the four compass directions. It is found that one feature produces a perfect match with another when the north of one is aligned with the east of the other. What does this tell you?

   • the second feature is a mirror image of the first
   • the second feature is rotated 180 degrees clockwise relative to the first
   • the second feature is rotated 90 degrees anticlockwise relative to the first

   • the second feature is rotated 90 degrees clockwise relative to the first

   Moving the north to the east is a 90-degree clockwise rotation, so this is the best match.

9. How are GLCMs normally used to identify similar textures?

   • some quantities are computed from them, with large values indicating a match
   • many values along the leading diagonal indicate similarity

   • some quantities are computed from them and they are matched

   • many values away from the leading diagonal indicate similarity

   A number of quantities are computed from the actual GLCMs, with the most common ones being listed in the lecture notes. These are then matched, most often these days using machine learning.

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

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

    • in the bottom of the dip between the two peaks

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

    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.