Abstract

1
A Generalized Linear Model for MPEG-2 Packet-Loss
Visibility Sandeep Kanumuri, Pamela C. Cosman and
Amy R. Reibman Department of ECE, Univ.
Calif. at San Diego ATT Labs Research
skanumur,pcosman_at_code.ucsd.edu amy_at_research.att
.com

• Null Model (Model 0)
• Model with only the constant parameter ? and
  with out any factors. This model predicts the
  same probability of visibility for all packet
  losses.
• Initial Model (Model 1)
• This model uses all the nine factors described
  in Table 1.
• Improved Motion Variables (Model 2)
• MOTM, HIGHMOT and VARM are introduced as factors
  in this model and factors MOTX, MOTY, VARMX,
  VARMY are removed.
• Improved Time-duration Variables (Model 3)
• FRAMETYPE is introduced as a factor instead of
  TMDR. FRAMETYPE informs us about the type of
  frame in which the packet loss occurred, whether
  I, B, P1, P2, P3 or P4. It is a categorical
  variable. This is our best model and final model.
• FRAMETYPE value for different frames in a GOP
• The following plot shows the reduction in
  deviance as we move
• from Model 0 to Model 3.

Abstract We focus on predicting the visibility
of packet losses in MPEG-2 compressed video
streams. We develop a generalized linear model
(GLM) to predict the probability that a packet
loss will be visible to an average viewer. The
GLM input consists of factors that can be easily
extracted from the video near the location of the
loss, and outputs an estimate of the probability
that that loss is visible. We also show how our
GLM can be used to classify each loss as visible
or invisible. Using this method, we are able to
achieve a high classification accuracy. Introduct
ion Packet losses are quite common in todays
communication networks and can have a detrimental
effect on the transmission quality of compressed
video such as MPEG-2. Network service providers
would like to (a) provision their network to keep
the packet loss rate below an acceptable level,
and (b) monitor the traffic on their network to
assure continued acceptable video quality.
Unfortunately, each packet loss in video has a
different visual impact, which makes the problem
of evaluating video quality given packet losses
very difficult and challenging. Our goal is to
develop a quality monitor that is accurate,
real-time, can operate on every compressed video
stream in the network, and answers the question,
How are the losses present in this particular
stream impacting its visual quality?. Towards
this goal, we develop a generalized linear model
(GLM) to predict the probability that a packet
loss will be visible to an average viewer. We
also show how our GLM can be used to classify
each loss as visible or invisible. Effect of a
packet loss We use zero error
concealment to conceal the lost packets. Zero
Error Concealment

• Subjective tests
• Task Viewers are shown videos with packet
  losses. They were asked to press the space bar
  whenever they see an artifact. An artifact is
  defined simply as a glitch or abnormality in the
  video.
• Single stimulus test
• 12 videos of 6 minutes duration each
• 3 videos in each set ? 4 sets of videos
• One set of video shown in each session
• Each packet loss evaluated by 12 viewers
• The videos we chose were from travel
  documentaries and are of DVD-quality with a
  resolution of 720 480 and 30fps. They were
  coded in the YUV 420 format. The compressed
  videos were coded at a high rate and did not have
  any coding artifacts.
• A total of 1080 packet losses were randomly
  injected in these videos such that every
  non-overlapping four-second interval contained
  one packet loss in the first three seconds. We
  distributed the losses such that 30 affected an
  entire frame, 10 affected two adjacent slices
  and 60 affected a single slice.
• Logistic Regression
• We model the probability of visibility using a
  Generalized Linear Model (GLM). Logistic
  Regression is a type of GLM which models the
  probability parameter p of a binomial
  distribution.
• Let y1, y2,, yN be a realization of independent
  random variables Y1,Y2,,YN such that Yi has
  binomial distribution with index mi and parameter
  pi. Let y, Y and p denote the N-dimensional
  vectors represented by yi, Yi and pi
  respectively. We are trying to model the
  parameter p as a function of P factors. Let X
  represent a N P matrix, where each row i
  contains the P factors influencing the
  corresponding parameter pi. A generalized linear
  model can be represented as

• Methods based on access
• FR Full Reference method
• RR Reduced Reference method
• NR-P No Reference Pixel based method
• NR-B No Reference Bitstream based method
• The above figure illustrates different methods
  for quality assessment based on locations for
  measuring video. In our work, we predict the
  probability of visibility using RR, NR-P and NR-B
  methods.
• Factors affecting visibility
• The following table describes the factors that we
  consider useful in predicting the probability of
  visibility.