Research Projects

Pablo Arbelaez










Hierarchical Boundaries


Original Image
Hierarchical Segmentation
Ultrametric Contour Map


We designed a boundary extraction system that combines several types of low-level image information into a generic notion of segmentation scale. Our system constructs a hierarchical representation of the image boundaries called Ultrametric Contour Map (UCM). Thresholding an UCM at level k provides by definition a set of closed curves, the boundaries of the segmentation at scale k. The system parameters were optimized with respect to ground-truth data, in order to make the UCM and the segmentation scale completely automatic.


Quantitative evaluation of Ultrametric Contour Maps on the Berkeley Segmentation Dataset and Benchmark (BSDB) using Precision-Recall curves.
  • Browse the results on the test set of the BSDB (100 images).

  • Take a look at my notes on the evaluation methodology.


  • P. Arbelaez. Boundary Extraction in Natural Images Using Ultrametric Contour Maps. Proceedings 5th IEEE Workshop on Perceptual Organization in Computer Vision (POCV'06). June 2006. New York, USA. [pdf]

  • P. Arbelaez. Une approche metrique pour la segmentation d'images. Ph.D. Dissertation. CEREMADE, Universite Paris-Dauphine. Paris, France. November 2005. (in French)