A flexible architecture for multi-view 3DTV based on uncalibrated cameras

Abstract

This paper presents a novel flexible architecture for 3DTV based on multiple uncalibrated cameras. The proposed signal representation improves the interactivity of dense point-based methods, making them appropriate for modeling the scene semantics and free-viewpoint 3DTV applications. The main concern is to address the shortcomings of depth image-based 3D video systems for free-viewpoint visualization, and to provide an efficient implementation of the rendering part which is computationally intensive as well potentially determine the view quality. Novel rendering algorithms are added that specifically aim at solving the rendering artifacts, and sampling issues encountered in wide baseline extensions and arbitrary camera movements. To optimize the process, a “selective” warping technique is proposed that takes the advantage of temporal coherence to reduce the computational overhead. Performance is illustrated on challenging videos to prove the suitability and flexibility of the architecture for advanced 3DTV systems.