Abstract

Global stereo approaches formulate the stereo matching problem in terms of an energy function that is subject to optimization. It is known that common energy functions represent a suboptimal model of the real problem so that even a global energy minimum leads to disparity maps of moderate quality. In this talk, I will present two papers recently published at CVPR that focus on improving the energy model.

The first paper concentrates on synergies that exist between the problems of stereo matching and alpha matting. We formulate both problems in a joint approach and show that our results are useful for the tasks of novel viewpoint synthesis and depth segmentation.

In the second paper, we model a stereo scene as a collection of a few 3D surfaces (planes and B-splines). We argue that this surface-based representation has several advantages over the commonly-used disparity-based representation, where each pixel is directly assigned to a discrete disparity value. In particular, it allows us (1) to incorporate the very popular color segmentation constraint in a soft way, (2) to formulate a global MDL prior, (3) to use a simple curvature term and (4) to improve occlusion handling. I will briefly mention an extension of this work (currently under review), where we use a similar framework to extract scene objects and model interactions between them.