Section: New Results

Stereoscopic vision

Current approaches to dense stereo matching estimate the disparity by maximizing its a posteriori probability, given the images and the prior probability distribution of the disparity function. This is done within a Markov random field model that makes tractable the computation of the joint probability of the disparity field. In this framework, we investigated the link between intensity-based stereo and contour-based stereo. In particular, we properly described surface-discontinuity contours for both piecewise planar objects and objects with smooth surfaces, and injected these contours into the probabilistic framework and the associated minimization methods. One drawback of such an approach, and of traditional stereo algorithms, is the use of the frontal parallel assumption that bias the results towards frontal parallel plane solution. To overcome this issue, we have investigated the use of a joint random Markov field, so that to each pixel is associated a disparity value and a surface normal. The estimation of the two field is done alternatively using minimization methods described above [56]