Hi, I am reading this toturial and try to make my own HOG-SVM hand detector: http://www.learnopencv.com/histogram-of-oriented-gradients/
I train some of the hand picture in libsvm, I resize the training data picture to 48x48 and use 12x12 cell to get the hog feature, then I have about 324 feature to represent the hand. After I get the model, I use sliding window to detect hand in image, my sliding window is 72x72 and cell size is 18x18 in order to fit the feature number to my model.
the result is not bad( accurancy ~= 80% ), however, I find that it would always get better accurancy if the stride is smaller than the cell size, for example, I can almost detect every hand (accurancy = 99% ) if the stride is 6 instead of 18, but 6 is not the smallest size if the cell size, and none of the document I read about HOG do such a thing. They always use the cell size as smallest stride. It would be difficult to make a whole HOG map of input image if the stride is smaller than the cell size ( I have to repeatly caculated hog window by window instead of caculated once and just find the ROI in general case)
what should I do if I find that the sliding window with smaller stride than the one in HOG descriptor would get the better result? should I just re-train my model with such a small cell (12x12) ? I'm afraid that if I do so the HOG feature would be much larger and the program would be really slow, but if I want to use the sliding window stride which is smaller than the cell size, It is already slow enough.
thanks for any advice!