Virtual Image Generation by View Extrapolation of Real Scenes

View synthesis involves use of existing images to generate virtual images at varying viewpoints. With the advancements in digital media, virtual view generation has gained much attention. It is widely being used in computer generated animations where it is required to transform one object into another, or to observe the same object from a different point of view. A variety of techniques exist, ranging from 3D modeling to 2D image-based rendering each offering a different set of advantages and limitations. In most of the work on view synthesis, the target viewpoints are usually restricted within the scope of real images and determined using interpolation. This paper describes a methodology to extend the concepts of interpolation to be able to generate new images at viewpoints that are not bound to be present within the virtual boundaries of reference views. Extrapolation has always been a prime method of predicting future states and behavior of elements and processes in all fields of technology. The focus of this paper is on the generation of virtual views using extrapolation, thus allowing surpassing the boundary that has always constrained the virtual viewpoints to lie within the scope of existing views. Special scenarios have also been identified, e.g. planar scenes, pure rotational camera motion etc. that allow faster view generation while avoiding the cumbersome and expensive computations. The paper is based on the PhD research of the author, whose thesis was successfully defended.