create multi-class labelling example with higher-order potentials

[amb-enthusiast]

This is where PyMaxflow might let us down, and you cython wizardry may help. If these methods accept data structures for pairwise potentials pixel labels (not just class label level) that would be great.

[jsherrah]

certainly I can put whatever edge weight we want on the graph, whether that achieves what we want for multi-class depends, I have to go study ab-swaps etc first. I think the graph cut approach requires that for label 0 the weight is zero, and for label 1 it's whatever cost you put on the edge.

[jsherrah]

do superpixels count as higher order potentials? Is that what this was about?

[amb-enthusiast]

My take is that we mean pairwise potentials over the superpixel network, rather than higher-order potentials over > 3 superpixel variables.

[jsherrah]

misnamed, we meant superpixels