Multiscale Image Segmentation

Ronen Basri
Weizmann Institute of Science

Minimal cut algorithms provide a popular means to achieve image segmentation through the partitioning of a graph induced by an image into regions (connected sub-graphs) of coherent color or texture. As multiscale measurements are clearly essential for achieving satisfactory segmentation results, an important question is how to acquire and use such measurements within the minimal cut framework. In this talk I will present our work on multiscale image segmentation by using principles adapted from Algebraic Multigrid methods. I will further discuss related methods for motion segmentation, automatic cue integration, and scale selection, and will present applications to detection and classification in the context of natural and medical imagery.

Presentation (PowerPoint File)

Back to Graph Cuts and Related Discrete or Continuous Optimization Problems