Hierarchical Pairwise Segmentation Using Dominant Sets and Anisotropic Diffusion Kernels
Pairwise data clustering techniques are gaining increasing popularity over traditional, feature-based central grouping techniques. These approaches have proved very powerful when applied to image-segmentation problems. However, they are mainly focused on extracting flat partitions of the data, thus missing out on the advantages of the inclusion constraints typical of hierarchical coarse-to-fine segmentations approaches very common when working directly on the image lattice. In this paper we present a pairwise hierarchical segmentation approach based on dominant sets  where an anisotropic diffusion kernel allows for a scale variation for the extraction of the segments, thus enforcing separations on strong boundaries at a high level of the hierarchy. Experimental results on the standard Berkeley database  show the effectiveness of the approach.