New view generation from a video sequence

Abstract

The problem of three-dimensional (3-D) scene representation and view generation is a new subject in video signal processing. Our goal is to construct a new scene at a different viewing position based on a sequence of captured images. The method proposed here does not make use of a solid 3-D model of physical objects. We decompose the image into layers based upon the motion maps derived from consecutive frames of an image sequence, and then synthesize a new view by recomposing the layers. In this process, accurate motion information estimation and consistent layer data aggregation are two critical steps. Some results using a real image sequence are shown at the end.