Compact real-time modeling of seated humans by video sprite sequence quantization

We propose an image-based method for real-time modeling of seated humans using upper-body video sprites, which is suitable for applications such as teleconferencing and distance learning. A database of representative video sprite sequences is pre-acquired and pre-uploaded to each remote rendering site. At run-time, for each input sprite, a closely matching sprite is located in the database and the index of the matching sprite is sent to the rendering site, which drastically reduces the data rate. Unlike other data compression methods, our method takes advantage of the limited number of significant body positions a participant assumes during a session. Exploiting redundancy between frames with distant time stamps enables aggressive compression rates with high visual and semantic fidelity.