First, as alluded to, typical video-encoded bitstreams are highly structured, e.g., characterized by a natural hierarchy of importance layers or resolutions. Examples of ``highly important'' encoded video signal descriptors include motion information used to reduce temporal redundancy, anchor-frame (e.g. MPEG I-frames) data, ``coarse'' information such as DC and low-frequency AC transform data etc. Examples of ``less critical'' descriptors include motion-prediction error high-frequency ``detail,'' etc. Secondly, the connection experiences non-negligible packet losses and consequent rate variations even if the network resources available for the connection remain invariant. This can lead to a potentially high variance in the delivered video quality.
We see that there is an inherent personality mismatch between the source coding algorithms -- that are priority-oriented or multiresolution (MR) in character -- and the network layer mechanisms in the Internet that are not endowed with the ``smarts'' to discern prioritized classes. This therefore underlines the need for an efficient ``transcoding'' mechanism that converts the scalable MR-based prioritized video bitstream into a non-prioritized one that is better matched to the Internet. We propose a novel way of doing this conversion based on Multiple Description (MD) coding principles, anchored on Forward Error Correction (FEC) channel codes. While MR-to-MD transcoding addresses the issue of the sensitivity of a packet flow to the relative position of packet losses, we also need to address the issue of fluctuating rate.
From the end user's point of view, a
multimedia application that delivers a medium but constant quality is
generally more desirable compared to one that provides quality that is high in
amplitude but exhibits large variations over time.
In this paper, we propose a congestion control algorithm that achieves
low rate fluctuations when the connection
capacity is invariant, as well as quick response to sudden changes in the
connection capacity (Linear Increase, Graded Multiplicative Decrease).
The novelty of our congestion control algorithm is that it achieves lower
fluctuation (hence higher throughput or average transmission rate) than
the state-of-the-art
congestion control algorithms without compromising the responsiveness to
the onset of congestion. The novelty of our approach lies in the
fact that
the source transcoding module and the congestion control module exchange
connection
state in a very simple manner, and yet the transcoder uses this information to
generate an MD stream that results in the optimal expected quality at the
receiver in spite of the dynamically fluctuating connection capacity. A
block diagram of the end-end streaming system follows.
