PROJECT TITLE :

One-View Occlusion Detection for Stereo Matching With a Fully Connected CRF Modelí_

ABSTRACT:

A belief propagation (BP) sequential method described in the tree-reweighted sequential method is extended to completely linked conditional random field models with the geodesic distance affinity in this study. The stereo matching problem has been solved using the proposed strategy. Another technique to the BP marginal solution, which we name one-view occlusion detection, is proposed (OVOD). The proposed OVOD approach enables for the discovery of occluded regions in the disparity map and the improvement of the matching result, unlike the usual winner takes all estimation. A single energy-minimization operation is all that is needed to avoid the second view's costs and a left-right check. For cost augmentation and energy reduction, the OVOD approach significantly outperforms the typical one-view affinity space implementation. With our strategy, we are able to achieve the best results for median, average and mean squared error metrics for the Middlebury data set.