Visual Correspondence Using Energy Minimization and Mutual Information

• Authors: Junhwan Kim;Vladimir Kolmogorov;Ramin Zabih
• Venue: ICCV '03 Proceedings of the Ninth IEEE International Conference on Computer Vision - Volume 2
• Year: 2003

We address visual correspondence problems without assuming that scene points have similar intensities in different views. This situation is common, usually due tonon-lambertian scenes or to differences between cameras. We use maximization of mutual information, apowerful technique for registering images that requiresno apriori model of the relationship between scene intensities in different views. However, it has provendifficult to use mutual information to compute densevisual correspondence. Comparing fixed-size windowsvia mutual information suffers from the well-knownproblems of fixed windows, namely poor performanceat discontinuities and in low-texture regions. In thispaper, we show how to compute visual correspondenceusing mutual information without suffering from theseproblems. Using a simple approximation, mutual information can be incorporated into the standard energyminimization framework used in early vision. The energy can then be efficiently minimized using graph cuts,which preserve discontinuities and handle low-textureregions. The resulting algorithm combines the accuratedisparity maps that come from graph cuts with the tolerance for intensity changes that comes from mutual information.