View synthesis from two uncalibrated images

Abstract

View synthesis becomes a focus of attention of both the computer graphics and computer vision communities. We present a new approach for synthesizing novel views from two uncalibrated images. The two reference images as well as the novel view do not share the same viewpoint. The developed approach incorporates computer vision methods. It consists of two stages. First, the parallax field between the reference images is recovered. Second, novel images are directly synthesized by exploiting the parallax invariance, using forward warping. Solutions to the visibility problem are proposed. Constructing realistic synthesized views from real image pairs are presented.