The existing code passes depth information through as one of the color channels, rather than handling it separately. However, the paper describes using histograms to build a model of the depth distribution, and combining it with the GMM color distribution for the edge weights.