Stereo Vision Correspondence Selection using Mutual Information and the Development of a Parametric Framework

This paper addresses the problem of correspondence selection in stereo vision using the mutual information (MI) measure. MI is an information theoretic topic that has recently seen a prolific expansion in the computer vision and medical imaging field. Two main issues are considered in this paper. Firstly, a new stereo matching algorithm is presented that uses a histogram-based formulation of mutual information. It also incorporates adaptive window sizes based on the amount of information contained in the candidate and template windows. Secondly, we present a discussion on some recent work on developing a parametric framework for MI stereo matching. This involves estimating parameters of the underlying probability density functions and then solving for the entropy of the estimated densities to compute the MI measure. The form used for the density functions are the generalised Laplace distribution and a Gaussian mixture model. Entropy calculations are then discussed using analytic methods and also a numerical Gauss-Hermite quadrature technique. Experimental results for the histogram-based stereo matching technique indicate that the mutual information measure shows significant promise due to its robustness in overcoming the effects of radiometric distortion. Preliminary results are also shown for the parametric method.