Transcript ContentBasedMPEGVideoTrafficModeling

Content Based MPEG Video Traffic
Modeling
Ali M. Dawood and Mohammed Ghanbari, senior member, IEEE
Presented by
Premchander Reddy & Lakshmi deepthi Pasupuleti
To
Donald Adjeroh
As a partial requirement for course CS558
What is video modeling?
Video model is an aid for designing and
testing future communication networks that
will carry multiplexed video traffic. It is an
essential tool in estimating many networking
issues such as the delay arising from
statistical multiplexing and the bandwidth
required for carrying video
Survey…..Classic Modeling
Non MPEG
MPEG
 Maglaris
 Sen
 Grunenfelder
 Heyman
 Hughes
 shim
 Pancha
 Heyman
 Wu
 Krunz
 Ni
Classic Modeling
• In the classical modeling the mean and variance of
real video are matched to an AR ( Auto regressive)
model or any known distribution function. The nature
of the video content and the length of the video is
not taken into consideration here.
• But modeling a video considering its nature and
content can obviously result in better representation
of video.
Introduction: Content Based Modeling
• Decomposition of video
Video Clip
Stories
Shots
GOP
Video Frames
: Such as a Film
: Such as News session
: Continuous action
: Group of Pictures
: I P B frames
Shot Classification
• Shot is a homogeneous video
• Modeling of video should
•
•
•
•
start from modeling of shot.
Texture and Motion are used
to classify shots into groups
3 levels of texture and 3
levels of motion are chosen
The levels are namely LL LM
LH ML MM MH HL HM HH
L M H stand for Low Medium
and High respectively
Measuring the Texture and Motion
• Texture: The average magnitudes of the
DCT coefficients of luminance/block for
each frame is calculated and then
averaged over the shot.
• Motion: The magnitude of motion
vectors/macro block are extracted for each
frame type and then averaged over the
shot.
• The relation between average
•
•
•
•
•
DCT coefficient and bit rate is
distinct for the I-frame.
Due to motion it is not so
distinct for P and B frames.
So we take the texture
information from the I-frames.
The motion information Is
taken from P and B frames
since I frame is intra-frame
coded.
I frames are combined with
those of P and B frames for a
reliable classification.
For example the classification
of texture can be known from I
frames and motion-based
classification is known from P
and B frames.
Characterization of Real Video
• The shot classification
•
•
•
were applied to a 30 min
BBC news bulletin.
The frequencies of
occurrence of each shot
type was tabulated.
The transition probability
table was also tabulated.
Transition probability
table gives us the
probability of a particular
shot type following the
next type.
Composition of Video Clips
• Mean bit rate is calculated for
each shot type and is divided
into I,P,B frames.
• After classification of shots and
determination of bit rate,
proportion of I,P,B bit rates,
the shot can be defined as a
vector.
Sk(AR_Ii, AR_Pi, AR_Bi, tk)
k=1,2..N is the kth shot in a clip of
N shots, i=1,2..9 is the ith shot
type,tk is the duration of the
kth shot.
Summary of synthetic generation
of CBM
• 1. Define the number of shots(N) in the video clip.
• 2. Specify the shot type and derive the mean bit rate of
•
•
•
each shot type, and derive the mean bit rate of each
shot from the overall mean bit rate.
3. Specify the shot duration, according to the statistics
and Gamma function.
4. Using the mean and variance, calculate the autoregressive (AR) model’s parameters for the kth shot[6].
5. Go to step 3 for the kth + 1 shot.
Results from Simulation of
Deterministic CBM
• The performance of the
•
•
proposed model was
tested against a real
video clip.
A virtual video clip was
edited from 11shots and
the proposed model was
applied.
It was observed that the
CBM traffic closely follows
the real non homogenous
MPEG traffic.
Realistic CBM
• Since the deterministic CBM is
•
•
•
based on subjective description of
the video content, the shot
classification may vary from
person to person.
In order to derive a more realistic
content based video model the
transition and the durations are
made probabilistic, based on the
shot characteristics.
A new shot type transition
probability table is formulated.
A nine-state model is used to
represent the probabilistic CBM.
Summary of Probabilistic CBM
• 1.Start from an initial state.
• 2. Find the duration of the slot with a gamma
•
•
•
•
function of α=2 and β=70.
3. According to the type of the shot, use the
table to calculate the auto-regressive (AR)
model’s parameters.
4. Run AR model for the duration of the shot
given in step 2.
5. Transit to the next state according to the new
transition table.
6. Go to step 2.
Comparison of Results
• A 2 min video clip was modeled with a classical AR
•
•
•
•
method, deterministic CBM and probabilistic CBM.
The worst performance was observed for the classical
method which do not consider video content.
The best performance was observed for deterministic
CBM .
The purely statistical probabilistic CBM had much better
performance than the classical model.
The network performance with these traffics was also
evaluated, where each model’s traffic has been fed into
an ATM multiplexer with network loads of 70% and
90%.
Network Performance
Limitations
1. Image representations based on low-level
2.
3.
visual primitives such as texture, and motion.
The determination of shots is also a complex
task.
Different people have different visual
perceptions, so classification of shots based on
color, texture and motion becomes a problem.
Suggested Improvements
• The classification of shots based on
contextual information such as appearance
of an anchor in a video can be useful.
• This type of classification is easy as the
contextual information from which the
classification is done is viewed as the
same by all the people.