Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Depth driven photometric and geometric image registration for real-time stereo systems

Tiefen-gesteuerte photometrische und geometrische Bildregistrierung für Echtzeit-Stereosysteme
: Waizenegger, W.; Feldmann, I.; Eisert, P.


Eisert, P.; Hornegger, J.; Polthier, K. ; European Association for Computer Graphics -EUROGRAPHICS-:
Vision, Modeling and Visualization, VMV 2011 : Berlin, Germany, October 04 - 06, 2011; Proceedings of the 16th Annual Workshop on Vision, Modeling and Visualization; Workshop chairs & proceedings
Goslar: Eurographics Association, 2011
ISBN: 978-3-905673-85-2
International Workshop on Vision, Modeling and Visualization (VMV) <16, 2011, Berlin>
Fraunhofer HHI ()

This paper presents a novel depth driven approach for a highly accurate joint photometric and geometric image alignment. Thereby, the registration problem is expressed in terms of a consistent, elegant and efficient energy formulation. Moreover, a real-time capable alternating iterative optimization scheme is proposed to solve for a state of minimal energy. Since the energy formulation is based on pixel wise color similarity the registration procedure directly improves the performance of disparity estimation and the visual quality of multi-view view synthesis. One goal is to optimize photometric camera settings with respect to optimal depth estimation results. Simultaneously it is aimed at the fully automatic on-line adjustment of colorimetric camera settings and at the electronic off-line fine-tuning of photometric properties for recorded stereo sequences. The work extends the variational formulation for disparity estimation, presented in literature formerly, with regard to improved color insensitivity and robustness. Beside off-line postprocessing of stereo images, it can be seen as a valuable algorithm for the on-line setup of stereo cameras used for disparity estimation. Simultaneously, as shown in the experimental results, the optimization procedure is directly linked to the performance of disparity estimation and the visual quality of multi-view view synthesis. The potential of the algorithmic development has been successfully demonstrated by the evaluation with two different applications and three different datasets.