This work will be focused at the creation and refinement of object class detectors with the aim of easing 3D reconstruction. Existing 3D reconstruction methods like shape-from-motion work in a strictly bottom-up fashion, in that sense that they are grounded in matching low-level information. As can be systematically observed, the resulting 3D quality is wanting. Either one arrives at a sparse, but precise reconstruction, or a dense surface representation is built, but with salient flaws. It would be of enormous help if such a 3D acquisition system could be made aware of the structures it is trying to model. The inclusion of high-level information is an important component. In particular, the plan is to let object or component class detectors influence the process. In his work, the candidate will focus on the timely issue of 3D building reconstruction. In particular, the role of detectors in the context of inverse procedural modeling will be investigated. Not only effective detectors for basic building components like doors, pillars, windows, etc. need to be built, they should also be invoked, combined, and refined at appropriate stages of the process.
Mathias M., ''Object detection for urban modeling'', Proefschrift voorgedragen tot het behalen van het doctoraat in de ingenieurswetenschappen, KU Leuven, October 2013, Leuven, Belgium.