In this paper we propose a real-time view synthesis implementation with valid geometry for a teleconference application with viewpoint adaptation capability, from which a wide range of realistic views can be reconstructed. The whole process is decomposed into two backward mapping steps and two 1D processing steps for enabling parallelism. We show that the proposed algorithm can provide high quality virtual views that are comparable with the real perceived view. Theoretical motivation and implementation issues of the view synthesis algorithm are discussed. Experiments show that, due to the decomposition, a real-time implementation is feasible. It is further argued that this view synthesis approach can be integrated seamlessly with other parts of the teleconference system, speeding up the virtual view reconstruction even more.
