Journal of Mathematical Imaging and Vision

, Volume 39, Issue 3, pp 230–244

Can Variational Models for Correspondence Problems Benefit from Upwind Discretisations?


DOI: 10.1007/s10851-010-0237-z

Cite this article as:
Breuß, M., Zimmer, H. & Weickert, J. J Math Imaging Vis (2011) 39: 230. doi:10.1007/s10851-010-0237-z


Optic flow and stereo reconstruction are important examples of correspondence problems in computer vision. Correspondence problems have been studied for almost 30 years, and energy-based methods such as variational approaches have become popular for solving this task. However, despite the long history of research in this field, only little attention has been paid to the numerical approximation of derivatives that naturally occur in variational approaches.

In this paper we show that strategies from hyperbolic numerics can lead to a significant quality gain in computational results. Starting from a basic formulation of correspondence problems, we take on a novel perspective on the mathematical model. Switching the roles of known and unknown with respect to image data and displacement field, we use the arising hyperbolic colour equation as a basis for a refined numerical approach. For its discretisation, we propose to use one-sided differences in the correct direction identified via a smooth predictor solution. The one-sided differences that are first-order accurate are blended with higher-order central schemes. Thereby the blending mechanism interpolates between the following two situations: The one-sided method is employed at image edges which often coincide with edges in the displacement field. In smooth image parts the higher-order scheme is used. We apply our new scheme to several prototypes of variational models for optic flow and stereo reconstruction, where we achieve significant qualitative improvements compared to standard discretisations.


Finite difference methods Adaptive methods Hyperbolic colour equation Variational correspondence problems Variational optic flow Variational stereo 

Copyright information

© Springer Science+Business Media, LLC 2010

Authors and Affiliations

  • Michael Breuß
    • 1
  • Henning Zimmer
    • 1
  • Joachim Weickert
    • 1
  1. 1.Mathematical Image Analysis Group, Faculty of Mathematics and Computer ScienceSaarland UniversitySaarbrückenGermany

Personalised recommendations