Section: New Results

Higher Order Graph Training throuh Dual Decomposition and Max Margin Principles

Paticipants: Nikos Komodakis, Bo Xiang, Nikos Paragios

In  [40] a novel framework based on the structure margin principle was introduced for training higher order graphical models. The idea was to reduce the training of a complex high-order MRF in the parallel training of a series of simple slave MRFs through a principled dual decomposition approach. The theoretical properties of the framework have been studied while the method has been experimentally tested using 2d/3d segmentation problems involving higher order geometric priors that are linear-invariant. The proposed formulation benefits from theoretical guarantees as it concerns performance, computational simplicity while being modular and scalable.