Wide baseline stereo based on background modeling

Abstract

Correspondence is one of the basic and most important problems in computer vision. Correct correspondence can be achieved through stereopsis and state of art algorithms have emerged for computation of dense correspondence incase of short baseline. Incase of wide baseline, maintaining correspondence in outdoor environment is a challenging task. First, A new concept of "motion from stereo" is presented, in which wide baseline is treated as summation of small baselines and correspondence is maintained all along it. To achieve this, correspondence based background modeling approach is extended to moving camera case. Second, occlusion among images is treated in a probabilistic manner and is modeled through Bayesian network. Third, a complete framework is presented to achieve correct correspondence. Experimental results validate that the framework is robust to nominal jitter, outdoor scenes artifacts e.g. swaying of trees, shimmering of water etc.