0

Lets say I want to train the cascade to recognize one object, but this object has different shapes. For example if I want to recognize a cup. You know there are cups in many shapes but also they have their similarities. Or another example: a steering wheel. You can always tell that the object is a steering wheel but it has different shapes. My question: Can I train the cascade to recognize most of the different shapes of one object?

dephinera
  • 3,703
  • 11
  • 41
  • 75

1 Answers1

2

if you train one cascade you will eventually get many false positives due to high variance of positive samples.

however if you have, say 20 stable classifiers, focused on different shapes, you can OR and use them.

suggestions on steering wheel case

  • do not use steered wheels as positives, only use 0 - degree wheels, this will make the classifier converge onto something.

  • group wheels by similarity in shape (color is not important for boosted cascade and background is not important because cars have similar gauge displays) and train cascades on these groups

  • search for a wheel in all rotations - note the symmetries - since boosted cascade is rotation invariant - and a steering wheel is likely to be steered.

baci
  • 2,528
  • 15
  • 28
  • thank you for your answer What do you mean by "OR" and also I'm not sure I understand you by saying to use only 0 degree wheels. – dephinera Nov 29 '14 at 21:19
  • logical OR. 0 degree = no steer. car moves straight :) do not supply rotations to classifier. – baci Nov 30 '14 at 04:08
  • So you say to create different classifiers for different shapes and use logical OR for the different xmls in my app when classifying the object? – dephinera Nov 30 '14 at 13:02
  • Okey, thank you! Could you suggest any parameters to use when training and also you can guess that there wont be a large number of positive images. Should I use create samples to create more positive samples? Which should be larger: the number of positive or the number of negative images? – dephinera Nov 30 '14 at 13:50
  • negatives should be larger than, at least 100 * positives. create samples would be a good enhancement too – baci Nov 30 '14 at 13:59