Distributed Video Coding

1
Distributed Video Coding

• Manzur Murshed
• Gippsland School of IT, Monash University
• Contributions from
• Mortuza Ali, Gippsland School of IT, Monash
  University
• Supported by
• ARC Discovery Project DP0666456
• International Postgraduate Research Scholarships
• Monash Graduate Scholarships

2
Outline

• Video coding
• Distributed video coding
• Distributed integer coding
• Integer coding with real domain predictions
• Coset-based prefix codes
• Existing Rice-Golomb prefix codes
• Performance analysis

3
Block-based Video Coding

• Key terms
• Macroblock
• Motion estimation and compensation
• Transformation
• Quantization
• Entropy coding
• Standards
• MPEG-1/2
• H.26X

4
Distributed Source Coding Slepian-Wolf Theorem
RX
RY
X
Y
Achievable rate region for two distributed
sources with identical distribution
5
Example Scenario I
X
X
ENCODER
DECODER
Y

• X and Y are 3 bit correlated data such that
  Hamming distance between them is at most 1
• Both encoder and decoder have access to side
  information Y
• Goal To encode X efficiently

Solution X?Y can take 4 distinct values
e.g. if Y 101 X can be 101,100,111,001
X?Y can be 000, 001, 010, 100 So, only 2
bits are required to encode X
6
Example Scenario II
X
X
ENCODER
DECODER
Y

• X and Y are 3 bit correlated data such that
  Hamming distance between them is at most 1
• Only decoder has access to side information Y
• Goal To encode X efficiently

• Solution The space of X is partitioned into 4
  cosets
• coset 0 000, 111 coset 1 001, 110
• coset 2 010, 101 coset 3 100, 011
• The encoder sends the index of the coset
  containing X instead of X
• On receiving the coset index of X, the decoder
  correctly recovers X from the coset which is
  closet to side information Y

7
Distributed Integer Coding
Side info
8
Encoder
Y 8
Y 8 X 7
Decoder
X Y ?max 1
1
C1, -2, 1, 4, 7, 10,
X 7
Encoder
7 is the nearest element to the side info
C0, -3, 0, 3, 6, 9, C1, -2, 1, 4, 7, 10,
C2, -1, 2, 5, 8, 11,
dmin 3 gt 2?max
Distributed Encoders
8
Encoder
Y 8
Y 8 X 8 1 7
Decoder
X Y 1 X Y ? -1, 0, 1
8
7 8 -1
X 7
Encoder
Conventional Difference Coding
8
Lossless Integer Coding with Real Domain
Predictions
Side info
7.8
Predictor
xt p, , xt 1
Decoder
xt 7
xt yt ?max 0.9
yt 7.8
1
C1, -1, 1, 3, 5, 7, 9,
xt 7
Encoder
7 is the nearest element to the side info
C0, -2, 0, 2, 4, 6, C1, -1, 1, 3, 5, 7,

dmin 2 gt 2?max 1.8
8
xt p, , xt 1
Predictor
Decoder
xt 8 1 7
xt yt 0.9 xt yt -1, 0, 1
yt round(7.8) 8
8 7 -1
Encoder
xt 7
Conventional difference coding incurs rounding
loss
9
Coset-based Prefix Codes
Side info
4.3
Predictor
xt p, , xt 1
Decoder
xt 7
xt yt ?t 2.7
yt 4.3
C1,a, -5, 1, 7, 13,
xt 7
Encoder
(i,j) (1,2)
7 is the nearest element to the side info
Given m 2 dmin m
C0, -2, 0, 2, 4, 6, C1, -1, 1, 3, 5, 7,

In order of nearest element to the prediction
C1,a, -5, 1, 7, 13, C1,b, -3, 3, 9, 15,
C1,c, -1, 5, 11, 17,
dmin (j1)m such that jm 2?t 5.6 lt
(j1)m i.e., j 2
Circled element is the only one within the range
10
Rice-Golomb Prefix Codes
4
Predictor
xt p, , xt 1
Decoder
xt 7
yt round(4.3) 4
xt yt ?t 3
M?t jm i 6 ?t xt yt 3 xt yt 3
7
xt 7
Encoder
(i,j) (0,3)
Given m 2 xt yt ?t 3 M?t if ?t ? 0
then 2?t else -2?t 1 6 j ?M?t/m? 3 i
M?t jm 0
11
Performance Analysis