Trifocal motion modeling for object-based video compression and manipulation

摘要: Following an overview of two-dimensional (2-D) parametric motion models commonly used in video manipulation and compression, we introduce trifocal transfer, which is image-based scene representation computer vision, as a compensation method that uses three frames at time to implicitly capture camera/scene depth. Trifocal transfer requires tensor computed by matching image features across views dense correspondence between two the views. We propose approximating model order apply for object-based compression background mosaic generation. Backward, forward, bidirectional methods based on are presented. The performance proposed approaches using has been compared with various other methods, such motion, block global affine transform several sequences. Finally, synthesis implemented within MPEG-4 Video Verification Model (VM), results those standard VM. Experimental show superior when there depth variation camera translation.