Automatic Selection of Reference Views for Image-Based Scene Representations

Abstract

The problem addressed in this paper is related to displaying a real 3-D scene from any viewpoint. To display a scene, a relatively sparse set of 2-D reference views is stored. The images that are inbetween the reference views are obtained by interpolation of coordinates and brightness (colour). This approach is able to generate the scene representation and render images automatically and efficiently even for complex scenes of 3-D objects. This is possible since the processing time does not depend on the complexity of the scene as there is no attempt to understand the semantics of images. In this paper we present a novel approach to automatically determine a minimal set of views from which the complete scene can be rendered. The method consists of two procedures: view-interval growing and selection. The first procedure independently searches for the intervals from which large portions of the scene can be rendered. These intervals are then passed to the selection procedure, which selects the minimal set of necessary views. The selection procedure is posed as an optimization problem that minimizes the number of reference views and the error due to the interpolation.