Haar classifier
What is the best size of weight and height to use in algorithm? Many tutorials use 24*24 pixels. My crop images haven't the same size but its more than 24 pixels.
Thanks in advance
Asked: 2013-11-06 16:01:19 -0600
Seen: 258 times
Last updated: Nov 06 '13
memory / train / detect time requirements grow exponentially with size, so try to make it not too large.
and maybe you should use opencv_traincascade, not the old opencv_haartraining (as it lets you try hog/lbp cascades as well)
ok thank you very much. I have 2 more questions. 1.Every tutorial that I found used opencv_haartraining. Can you suggest me sites that refers to opencv_traincascade? 2. Positive and negative images have to be the same size? Cause the region of interest that I want to detect has not the same size every time. Usually it is ~100*100 pixels.
I always suggest people with a general memory setup (lik 4GB of RAM) to reduce the largest dimensions in the region of 0 - 50 pixels and resize the smallest dimension accordingly. This will ensure that training can be performed completely. For example, if you have 500 x 100 pixel training images, take model size 50x10 which will resize all images first, but will guarantee a successfull training.