Résumé | This talk introduces our new, sub-pixel precise GeoMap as a unified
representation for both topological and geometrical aspects of a partition of
the plane (e.g. a 2D segmentation result). Our GeoMap is conceptually very
similar to the toger framework developed at the LaBRI, but is not based on
discrete maps using inter-pixel boundaries, but on a polygonal
representation.
This has several advantages: For image segmentation, it becomes possible to
combine the advantages of sub-pixel accuracy (as provided e.g. from the
well-known edge-detector by Canny) and topologically useful, closed
boundaries (as known from the traditional watershed transform), and lets us
introduce the exact watershed transform as the basis for hierarchical
segmentation approaches.
Furthermore, we demonstrate that our new sub-pixel GeoMap is also suitable for
applications beyond image segmentation, e.g. triangulation or
skeletonization. |