CVPR2006_Boundary Extraction in Natural Images Using Ultrametric Contour Maps英文文献资料.pdfVIP

Boundary Extraction in Natural Images Using Ultrametric Contour Maps ´ Pablo Arbelaez ´ CEREMADE, UMR CNRS 7534 Universite Paris Dauphine, ´ Place du marechal de Lattre de Tassigny, 75775 Paris cedex 16, France arbelaez@ceremade.dauphine.fr Abstract This paper presents a low-level system for boundary ex- traction and segmentation of natural images and the eval- uation of its performance. We study the problem in the framework of hierarchical classification, where the geomet- ric structure of an image can be represented by an ultra- metric contour map, the soft boundary image associated to Figure 1. Natural image and composite human segmentation. a family of nested segmentations. We define generic ultra- metric distances by integrating local contour cues along the regions boundaries and combining this information with re- of regions and scale are formalized by the structure of in- gion attributes. Then, we evaluate quantitatively our re- dexed hierarchy or, equivalently, by the definition of an ul- sults with respect to ground-truth segmentation data, prov- trametric distance on the image domain. We present a for- ing that our system outperforms significantly two widely mulation of the problem in terms of contours that allows to used hierarchical segmentation techniques, as wel