The g Factor: Relating Distributions on Features to Distributions on Images

The most common approach is to learn histograms of filter responses or, equivalently, to learn probability distributions on features (see right panel of figure (1)). See, for example, [6], [5], [4].