I modified train_HOG.cpp in OpenCV samples:
- training image size can be indicated with "-W" and "-H" options.
- hard negative training can be run with "-v" option.
- input format of positive sample text is the same as one of opencv_traincascade.exe.