The talk addresses two different topics of continuous and discrete optimization.
First, an extension of the ROF-model to vector-valued data (image
flows) is considered. The resulting regularizer differs from the commonly used term and gives rise to a decomposition into piecewise harmonic flow structure as opposed to piecewise constant flows, and motion texture.
The second part of the talk considers the MAP-inference problem and a QP-relaxation recently suggested by Ravikumar and Lafferty. As an alternative to rectifying the spectrum of the corresponding nonconvex quadratic forms, we study a class of convergent inference algorithms based on DC-programming. Experiments indicate that this alternative can be beneficial for highly-connected graphs with large state-spaces.
Christoph Schnoerr, University of Heidelberg Joint work with Jing Yuan (part 1) and Joerg Kappes (part 2)