Revision history [back]

I think the model histogram can be interpreted as a look up table of the distribution prob("pixel with color h,s shows skin") (if you normalize the histogram). Using this look up table allows you to have arbitrarily shaped distributions. In the case of skin colour, you probably could also fit a gaussian to the model histogram (using the number of pixels that fall into one hs-bin as weight) and then use this to compute the skin-probability.