Understanding traincascade code

edit

Thank you for your valuable suggestions. However, in fact, I both searched and read most of the sources what you recommended. I also read the Viola and Jones's paper about this. But, I still haven't been able to comprehend the code exactly. I have read http://docs.opencv.org/modules/objdetect/doc/cascade_classification.html. But in the opencv/apps/traincascade, some of the methods are not the same with this doc. The CvCascadeClassifier class in the traincascade folder is different from the class presented in the link. Where can I find the descriptions about all the methods in the code under opencv/apps/traincascade folder?

Thanks in advance.

Edit:1. The training data is prepared by fillPassedSamples and only neg. images are scaled. 2.The window size is equal to pos.image size. ??

To your edit: correct, since negative images can be of any size, and subwindows are used, whereas the positives have been resized to a data vector of fixed w and h by the createsamples utility and thus don't need extra scaling anymore. Most of the processing also goes to retrieving negative images that are classified as positives by the t-1 detector model that has been produced. So the later the stage of training, the longer it will take. About the algorithm itself, the problem with all machine learning is that it uses the old C api internally but with a C++ compatible skin around it. This makes indeed using the guidelines in the function calls dificult to match to the code. Best way to go, build the library yourself then start debugging and go into every single function :)

It would be wonderful if there were more comments for functions! Anyway, thanks for the code, of course.. Thank you.

Well yeah, since it is open source, OpenCV doesn't require tons of comments in code, expecially not in the past. Guess we have to live with that. However, if you would place explanations in code like this, feel free to submit it with a pull request.