Using Belief Propagation for Local Symmetry Detection in Images
Silke Jansen
Max-Planck-Institut für Informatik - D4
Talk
I received my bachelors degree in computational biology from Saarland University in 2008. At the same time I was accepted at the Saarbruecken Graduate School Computer Science. Currently I'm working on my masters thesis at the Statistical Geometry Processing group. I plan to graduate this Spring.
During the last decade, belief propagation (BP) has become widely used for inference on graphical models. Although not exact on general graphs, loopy belief propagation (LBP) often delivers surprisingly good results. The aim of this thesis is to retrieve reoccurring parts in images without any prior knowledge. Instead of heuristically extracting symmetries, a probabilistic model describing partial image matching is used. The marginals of this probability distribution are approximated using LBP. Finally, the desired symmetric parts can be extracted from the marginal distribution. After a short introduction to BP/LBP, the talk will cover issues that have to be considered when constructing the graphical model. Followed by a discussion on the expected quality of the results and variations of BP that promise to yield better results.