Text this: Towards closing the energy gap between HOG and CNN features for embedded vision