Generalizing the virtual camera pose for view synthesis

Abstract

View synthesis requires the ability to estimate the image projected from a scene to a point of view where has not been placed a real camera. Many methods have been developed and three-view rectification is one of the most used, nevertheless it has some restrictions. That arises when the plane containing the focus of the three cameras involved in the process is parallel to the view direction of any of the cameras. This paper deals with the geometry of the method and gives analytically a way for dodging the singularities in the position of the virtual camera. That allows us to obtain a synthetic view from a previously forbidden point and automate the process towards fast software or hardware implementations. © Springer-Verlag 2003.