<title>Scalable video coding over heterogeneous networks</title>

Abstract

Heterogeneous networks mix wired and wireless technologies, exhibiting well over three orders of magnitude of throughputs. Mobility adds the challenge of dynamic abrupt changes in the throughput experienced by a user. We describe in this paper a video service architecture that accommodates these widely different operating environments while providing graceful degradation as conditions change dynamically for mobile users. The video service is designed in the context of CMU's wireless Andrew. Key to our architecture is Generative Video (GV), a meta representation for video sequences. GV reduces video to still images, the world images, and ancillary data. The world images integrate the information about the background and about all moving objects present in the sequence. World images `lossless' compress the video sequence by at least two orders of magnitude. When coupled with spatial compression of these still images, the total compression ratio can easily exceed three orders of magnitude, well beyond the possibilities of current technologies. Scalable encoding of GV reduces to scalable encoding of still images. We discuss three alternatives, frequency scalability, signal to noise ratio scalability, and spatial scalability. We present results with a real video clip of 20 seconds duration, illustrating compression ratios of forty five hundred with acceptable (VHS type) quality. Due to the different nature and role played by the two types of GV data, we choose two transport protocols: TCP/IP for the transmission of the world images, and UDP/IP for the transport of the ancillary data. The paper describes the video system architecture and several design choice and optimizations that improve the overall performance of the video service.