Dear OpenCV Team and Users,
I wonder how positives are evaluated by opencv_traincascade. Is it required that the cascade returns a hit on each exact positive image, or is it sufficient to hit a subwindow of each positive image? In the latter case: are there parameters to control the maximum deviation, e.g. in scale?
I ask because I have the effect that my trained cascade frequently returns a hit on parts of the objects I want to detect, but not on the entire object.
Can someone point me to the code where the comparison of positives is done during training?
Thanks & best regards,
Michael