For the second upgrade, I decided to try creating a gradient value for histograms of oriented gradients.
Such histograms have seen widespread success, namely HOG and SIFT. It seemed logical to give them a try
in this context.
All I did was vector quantize the angle of the gradient at each pixel and then feed that vector quantization
through the same code that returns bg and tg, yielding a new value, gg.
Both SIFT and HOG use fancy binning and normalization techniques, none of which were employed here. I wanted to
see what I could get cheaply using the machinery at hand.
Original:
bg Mean:
tg Mean:
gg Mean:
Performance, 20 Images:
Performance, 200 Images:
