Rate-Distortion
Hint Tracks for Adaptive Video Streaming
- Distortion Chains
Predicting the distortion of a reconstructed media
presentation at a receiver in the event of missing (whether late or lost)
packets is difficult because of the complex distortion interactions associated
with the missing packets. We propose a mathematical model, which we call
Distortion Chains, for predicting the mean-square error (MSE) distortion at
the receiver in the event of packet loss (note that this also includes late
loss, i.e., media packets arriving after their delivery deadlines). This model
provides a simple, causal approach for predicting the distortion in the
reconstructed media presentation for general packet loss patterns. The idea
behing Distortion Chains is analogous to the Markov Chains model from
Probability Theory. Specifically, let k =
(k1,k2,k3,...,
kN,kN+1,...kM)
denote an arbitrary packet loss pattern of length M. Then, the contribution to
the overall distortion associated with the loss of packet
kn (for 1 ≤ n ≤ M), given that packets
k1,k2,...,kn-1 are
already lost, is a function of at most the P previous lost packets (see Figure
1), where P is the order of the Distortion Chain. Note that through P the
prediction accuracy and the complexity of the model can be traded-off. We
employ Distortion Chains to construct the distortion-rate tables for hint
track based streaming systems, as discussed next.
Figure 1. Distortion Chains for distortion
prediction.
- Low-complexity Rate-Distortion
Optimized (RaDiO) Streaming
The following research can be thought of as an intermediate
step between a streaming system that is completely ignorant of the importance
of individual packets and their interdependence and a streaming system running
our advanced RaDiO framework.
The motivation for such a research lies in the
fact that virtually all streaming systems existing in practice today can be
accurately described as the former one. Therefore, we feel that a streaming
system based on the research proposed in this section can be the first example
of intelligent or media aware streaming systems and thus can represent a first
step towards complete RaDiO streaming. In essence, such systems can motivate
the transition towards RaDiO streaming in the Internet and hence can make it
happen sooner.
We propose to design a streaming system based on a
rate-distortion table and simple algorithms. Specifically, the table contains
distortion-rate pairs (Di,Ri) for different
packet loss patterns, where Di represents the
reconstruction distortion associated with the loss pattern i, while
Ri is the size of the pattern in bits. The algorithms
comprise a method for predicting the distortion for a general packet loss
pattern based on the already existing entries in the table and a decision
method for finding the least harmful packet loss pattern in a rate-distortion
sense. We jointly denote the table and the associated algorithms as a hint
track as they are supposed to provide hints to a streaming system
regarding its transmission decisions. In particular, a streaming system can
then follow a hint track and make transmission decisions in the event of
insufficient transmission bandwidth and/or packet loss. An illustration of the
proposed system is shown in Figure 2.
Figure 2. Low-complexity RaDiO
streaming.
A hint track based streaming system achieves substantial
improvement in performance over a streaming system that is oblivious to the
distortion information associated with each packet, when determining the
transmission schedule for the video packets. Furthermore, these gains are
achieved with complexity comparable to that of the oblivious system. An
alternative approach to achieving low complexity RaDiO streaming is by
reducing the computational complexity of our advanced framework for packet
scheduling. As part of this study, we also investigated the complexity
savings relative to the advanced RaDiO framework and the associated
performance-complexity trade-offs.
A major part of this research was a joint work with the
Streaming Media Systems group from Hewlett-Packard Laboratories, in Palo Alto,
CA, where I spent the summer of 2003 as a visiting researcher.
|