Global Solutions for Shape Optimization in Computer Vision

Daniel Cremers
University of Bonn

Graph cuts allow to compute globally optimal solutions to a number of Computer Vision problems such as image segmentation, 3D reconstruction and planar shape matching. In my presentation, I will discuss strength and weaknesses of the graph cut method and relate it to alternative global optimization methods. Firstly I will focus on the cases of image segmentation and multiview reconstruction where the binary labelling problems can be optimized globally in a spatially continuous setting by reverting to convex relaxations. I will discuss advantages of this continuous analogue of the graph cut solution in the light of metrication errors, memory requirements and efficient parallel implementations. Secondly, I will focus on the case of planar shape matching and compare a recent graph cut formulation to a recursive shortest path formulation. Thirdly, I will show that the combinatorial problem of elastically matching shapes to images can be solved optimally by shortest path techniques in appropriate higher-dimensional spaces, thereby allowing to impose elastic shape priors in a globally optimal segmentation method.



*This is joint work with Kalin Kolev, Maria Klodt, Thomas Brox, Selim Esedoglu, Thomas Schoenemann, Frank Schmidt, Dirk Farin and Yuri Boykov.


Back to Long Programs