Professor Dr. Daniel Cremers
Inst. für Informatik
Universität Bonn

Continuous versus Discrete Optimization Methods in ComputerVision

A multitude of computer vision challenges can be cast as problems of energy minimization. I will introduce optimization methods which address problems such as the segmentation of moving objects in image sequences, the detection of obstacles in traffic videos, or the reconstruction of 3D shapes from a collection of 2D images. I will detail how respective cost functionals can be minimized both by continuous (PDE and level set methods) and by discrete (graph theoretic) methods.

In particular, I will show that the classical problem of multiview reconstruction can be solved by continuous convex relaxation methods. More specifically I will propose an integration of stereo and silhouette information as a convex stereo-weighted minimal surface problem over the convex domain of silhouette consistent solutions.