hog+svm: the size of positive image must be 64*128? any other choice?

edit

The sample size of a classifier has to be carefully chosen accordingly to the problem. For instance, the 64x128 size is commonly used in pedestrian detection, since a pedestrian image has obviously more height then width. Having that said, 64x128 isn't a good size for detecting your object as its aspect ration is wider. You must create a training set with your chosen sample size, to then perform detection with that same size.

You could even try 128x64 at first. It would be easier to implement because it is very similar to the widely used 64x128, and then if performance is not good enough, try to tweak it until you are satisfied