The current version of opencv train cascades seems to be single threaded other than precalculation steps and very very slow.
I'm wondering if:
a) there's any GPU accelerated version of opencv_traincascade.
b) if there isn't, is there a documented version or some pseudocode for it, I could try to parallelize it myself.
c) is using HAAR cascade worth it? Do they even perform well in non-faces? I'm using this for street signs and differentiating between different types and there's a lot of incorrect classification.
d) if it's not worth it, what would be better for classifying images and more importantly, searching for a subset of the image that matches it.
e) general HAAR cascade question: why does the cascade provided with opencv for faces capture the entire face while one I generate only gets the size of the cascade? I checked and the cascade provided is only 24x24. Why does detect multi scale show the right scales for only the default face detection system?
↧
