Parametric view-synthesis

Abstract

We present a simple procedure for synthesising novel views, using two or more basis-images as input. It is possible for the user to interactively adjust the viewpoint, and for the corresponding image to be computed and rendered in real-time. Rather than employing a 3D model, our method is based on the linear relations which exist between images taken with an affne camera. We show how the `combination of views' proposed by Ullman and Basri [19] can be appropriately parameterised when a sequence of five or more images is available. This is achieved by fitting polynomial models to the coeffcients of the combination, where the latter are functions of the (unknown) camera parameters. We discuss an alternative approach, direct image-interpolation, and argue that our method is preferable when there is a large difference in orientation between the original gaze directions. We show the results of applying the parameterisation to a fixating camera, using both simulated and real input. Our observations are relevant to several applications, including visualisation, animation, and low-bandwidth communication.