A new prediction structure for multiview video coding

A new prediction structure for coding multi-view video streams is presented. In general, for Free viewpoint TV (FTV) applications, it is necessary that multi-view videos are efficiently compressed before transmission. Our algorithm synthesizes extra video streams and uses them as extra references when coding the original views. These streams are synthesized based on the already encoded frames from neighboring views, without requiring the scene's depth information. The proposed scheme utilizes both motion and disparity compensation methods to exploit temporal and inter-view correlation within each view sequence and among views, respectively. To guarantee the best bitrate performance, our algorithm adaptively re-sorts the reference frame list, such that minimum number of bits is used for coding reference frame indices. Performance evaluations show that our proposed coding method outperforms the recent multiview coding standard by up to 1 dB PSNR and enhances the compression ratio by 22.97%.