VR models from epipolar images: An approach to minimize errors in synthesized images

Abstract

A new paradigm, the minimization of errors in synthesized images, is introduced to organically combine Computer Vision and Computer Graphics for Virtual Reality applications. Based on it, a powerful algorithm, called the strip DP algorithm, is proposed for epipolar image analysis. The algorithm reconstructs VR models from epipolar images so that the error in the images synthesized from the extracted model is minimized while geometrical consistency is maintained. The dynamic programming technique, adopted as the optimization engine, yields complete optimization at reasonable computation cost in a robust way. The strip DP algorithm is a multi-pass solution to occlusion problems, and, in each pass, it extracts connecting feature lines that are not occluded by undetermined feature lines. Experiments demonstrate its feasibility.