Hierarchical Segment Support for Categorical Image Labeling

This paper introduces a novel method for categorical image labeling, where each pixel is uniquely assigned to one of a set of unordered, discrete labels. Starting from provided label-depending pixel likelihoods we (a) exploit a segment hierarchy as spatial support to define powerful priors and (b) introduce an efficient and effective inference method, that can be implemented in a few lines of code. Experiments show that competitive labeling accuracy compared to related discrete, continuous, segmentation and filtering approaches is achieved.