Forward error correction (FEC) based schemes are used widely to address packet loss problems for Internet video. Given the total available bandwidth, finding the optimal bit allocation is very important in FEC-based video, because the FEC bit rate is deducted from the total rate. We want to give proper protection to the source, but we also should prevent unwanted FEC rate expansion. The rate of packet headers is often ignored in allocating bit rate. Actually, this packetization overhead has significant influence on system performance in many cases. Reducing the packet size will increase the rate of packet headers, thus reducing the available rate for the source and its FEC codes. On the other hand, smaller packet size will allow a larger number of packets, in which case it can be shown the efficiency of FEC codes improves.
This gives rise to several important questions: what exactly is the tradeoff between the packetization overhead and the forward error correction? How important is the optimal point (i.e. how much gain is there in achieving optimality, as opposed to an ad-hoc approach)? Finally, if optimality is indeed practically important, how can one achieve the optimal point?
We solve this problem in the context of one of the important video coding modalities for networks: scalable video. Past work in FEC-protected scalable video for networks has ignored the effect of packet overhead, or at most has minimizing overhead without considering the effect on efficiency of FEC. We show that significant gains are possible, without much pain, by a judicious allocation of rate while keeping the overhead in mind. We implement our solution on the MPEG-4 Fine Granularity Scalability (FGS) mode. FGS, a very simple and flexible layered coding method, has a two-layer structure where the enhancement layer is progressively coded. To show the flexibility of our technique, we use an unequal error protection scheme with FGS, and present an overall solution for packet size, code rate for the base layer, and code rate for the enhancement layer.
Setup and System ModelThe FGS mode of MPEG-4 is especially useful in our demonstrations because it has both progressive and a non-progressive parts, thus allowing us to construct fairly general models and experiments. Our approach can be generalized easily to bitstreams consisting of several progressive and several non-progressive components. The packetization of video and corresponding rates are shown in the figure below. The area of the entire group of packets is fixed, representing the overall rate constraint. Subject to this rate constraint, the packet size, as well as the proportion of enhancement layer bits, and parity bits for all components, are all determined by our algorithm without any further constraints.
We represent the network channel with a two-state Gilbert-Elliot model, a common model for problems of this sort that has been experimentally verified. Using a Lagrangian formulation, we can solve for the bit allocation problem in the presence of overhead.
Experimental ResultsThe formulation of the problem and its solution are available in reference  below. We reproduce two of the plots here. The first shows PSNR values as a function of channel conditions, parameterized by the packet size. It shows that optimal packet sizing can improve overall PSNR by 1-2 dB. The second plot shows the profile of optimal packet sizes, as determined by our algorithm, as a function of channel conditions.
Acknowledgments:This research was made possible in part by the National Science Foundation through grant CCR-9985171, and in part by a grant from Texas Higher Education Coordinating Board.
Last modified April 2002
Back to MCL Main Page