What is the best way to train a cascade classifier to find this sort of artificial pattern?
I tried running several training sessions with images from googlecode's haartraining tutorial varying from 500 negatives : 1000 positives to 1500 negatives : 4000 positives. Minimum false alarm rates varied between 0.4 to 0.5 and minimum hit rate varied from 0.995 to 0.999. Several tests were conducted with 10-15 stages, each resulting in an acceptance ratio of <0.001 at the very last stage.
I'm assuming my trainingdata is not good, because with each XML file the cascade classifier was not able to detect the macbeth chart. So I want to ask the more experienced people here how they would set up training for this pattern?