Discrete MRFs in computer vision and medical imaging: Deformable image fusion & 3D reconstruction with shape grammars

N. Paragios, Ecole Centrale de Paris, INRIA, France

Wednesday, November 4
at 12.15, HG D 5.2

Computer aided human perception aims at developing intelligent algorithms towards understanding visual cues coming from images, video, or other means of gathering visual information. Such a process often consists of three stages, initially the problem of perception is parameterized through a mathematical model where the estimation of its parameters will lead to visual understanding. Then, the model is associated with the available observations through the definition of an objective function and last, this function is optimized using computational methods. The main challenges that one has to address in this context is the curses of dimensionality, non-linearity, non-convexity and modularity. In simple words, even the simplest possible perception problem could involve too many parameters where the association between the data and them is not straightforward and is done through non-convex functions. In this talk, we will present a generic mathematical framework that exploits recent advances in discrete optimization to address computational visual perception. In particular we will present a class of algorithms that do exploit efficient linear programming, duality for low and higher order models. The challenging problems of inter/intra-modal deformable registration and 3D reconstruction using shape grammars will be considered to demonstrate the potentials of this framework.