Interactive Labeling of Image Segmentation Hierarchies

We study the task of interactive semantic labeling of a given segmentation hierarchy andpresent a framework consisting of two parts: an automatic component, based on a ConditionalRandom Field whose dependencies are de ned by the inclusion tree of the segmentationhierarchy, and a feedback-loop provided by a human user. Experiments on two data setsshow higher classi cation rates for the proposed framework than a baseline, that classi esall segments independently.