A quiz on the broken biscuits experiment

Adrian F. Clark

1. How are biscuits distinguished from the background in bikky.py?

   • by shape

   • by grey-scale

   • by colour in HSV space
   • by colour in RGB space

   The program segments foreground from background by thresholding. It might work better if we were to identify the background by colour and then invert the mask...

2. How is a biscuit determined to be broken or unusable?

   • it is not circular or rectangular

   • it has a large rectangularity
   • it has a large circularity
   • it cannot be segmented

   There is no single way to do this but pretty much everyone identifies biscuits that are unbroken, and hence circular or rectangular. Other ones are then broken or unusable.

3. What is a sensible threshold on the rectangularity for a biscuit to be identified as rectangular?

   • 1
   • about $4\pi$
   • 0

   • about 1.5

   Theory suggests that a threshold of about 1 should work but in practice it needs to be larger. About 1.5 works reasonably well.

4. How many images are used for training in the task file?

   • 6

   • 219
   • 213
   • 100

   There are only 6 files for training specified in the task file.

5. What is a sensible threshold on the circularity for a biscuit to be identified as circular?

   • about 1
   • $4\pi$

   • about 15

   • about 12.5

   The smallest possible value is $4\pi$ but the segmenter rarely produces a perfectly circular contour. Experience suggests that a limit of about 15 works on this dataset.

6. How do you make bikky.py iterate over the test set in biscuits.task?

   • edit the task file
   • edit the program
   • use -test on the command line

   • use -dataset test on the command-line

   The ability to work on different datasets in the task file is built into bikky.py: you use the -dataset qualifier with the name of the dataset to process.

7. How many images contain broken or unusable biscuits?

   • 100

   • 22

   • 219
   • 6

   There are 22 images in the task file which are identified as "rej".

8. What contour measure must be found for a rectangular biscuit?

   • axis-aligned bounding box

   • oriented bounding box

   • convex hull
   • minimum enclosing circle

   It makes sense to use the bounding box which is aligned with the edges of the biscuit rather than the edges of the image.

9. What is the purpose of this experiment?

   • to see how a complete vision system works on real-world data

   • to identify which biscuits are broken
   • to see how to segment images by colour
   • to use ways of describing shape

   Although all the answers are true to some extent, the REAL reason for this experiment is for you to learn that techniques which look as though they should work (like using circularity to identify circular regions) are normally much less neat than the theory would have you believe.

10. What type of problem is this defined as in the task file?

    • segmentation
    • biscuit

    • label

    • classification

    The task file defines the problem as a 'label' one, meaning there is a single label to be determined for each image.