View Interpolation for Dynamic Scenes

Abstract

This paper describes a novel technique for synthesizing a dynamic scene from two images without the use of a 3D model. A scene containing rigid or non-rigid objects, in which each object can move in any orientation or direction, is considered. It is shown that such a scene can be converted into several equivalent static scenes, where each scene only includes one rigid object. Our method can generate a series of continuous and realistic intermediate views from only two reference images without 3D knowledge. The procedure consists of three main steps: segmentation, morphing and postwarping. The key frames are first segmented into several layers. Each layer can be realistically morphed after determining its fundamental matrix. Based on the decomposition of 3D rotation matrix, an optimized and unique postwarping path is automatically determined by the least distortion method and boundary connection constraint. Finally, four experiments, which include morphing of a rotating rigid object in presence of occlusion and morphing of non-rigid objects (human), are demonstrated.