Title: DistortionAware Retransmission of Video Packets and Error Concealment using Thumbnail
1Distortion-Aware Retransmission of Video Packets
and Error Concealment using Thumbnail
- EE398 Course Project
- Winter 07/08
- Presenter Zhi Li
2Presentation Outline
- Idea in a nutshell
- Realizations
- Distortion estimation
- Retransmission decision
- Adaptive error concealment
- Experiment results
- Conclusions
3Retransmission-Based Robust Video Streaming over
Lossy Network
0101
1101
0001
0100
NACK
NACK
1101
0001
What if retransmission is subjected to a rate
constraint?
4Prior Related Work
- Soft ARQ Vetterli98
- avoids retransmitting packets that already passed
deadline - Syntax-aware (frame-aware) Zakhor02
- gives retransmission priority to I packets over P
and B packets - Analysis-by-synthesis De Martin07
- With NACKs, Sender simulates decoding (with ErC)
multiple times and decides retransmission based
on distortion
5Proposed Approach Thumbnail-Aided Retransmission
and Error Concealment
0101
1101
0001
0100
NACK
NACK
1101
0001
6System Overview
Raw video
Video Packets
Lossy video
Video Encoder
Lossy Network
Video Decoder w/ErC
Locate Error Packets
Send NACK
Thumb. Generator
Thumbnail
7Decomposition of Module
Raw video
Video Packets
Lossy video
Video Encoder
Lossy Network
Video Decoder w/ErC
Locate Error Packets
Send NACK
Thumb. Generator
Thumbnail
Lossy video
Slice Dist. Estimation
Pkt Dist. Estimation
Retrans. Decision
Thumbnail
8Thumbnail Generation Slice Distortion
Estimation
- Two types of projections
- Projection into the mean
- Random projection (J240)
Z
X
Y
Block
Projection
Projection
Thumbnail pixel
Welsh- Hardamard Transform (WHT)
Est. MSE
e.g. 4 bits
9Comparing Two Projections
Projection into the mean
J240 random projection
(Foreman CIF)
10Packet Distortion Estimation
- Distortion model decaying and additive
- Knowing slicing distortion, we can solve
distortion contribution of lost packets through a
set of linear equations
11Packet Distortion Estimation (Contd)
(Foreman CIF)
12Retransmission Decision
.
.
TR
- Retransmission priority based on
(assume packet size known) - Sender randomly drops B packets to maintain 100
transmission rate (also compensating thumbnail
rate)
13Adaptive Error Concealment
Control Data
DCT Coefs
Decoder
Deq./Inv. Transform
0
Motion- Compensated Predictor
ErC
Intra/Inter
Motion Data
Thumbnail
14Comparing Various Schemes
- Oracle assume receiver knows orginal video
- Thumbnail-aided proposed method using mean-based
thumbnail - Frame-aware receiver requests retransmission
based on the assumption that I packets are more
important than P packets - No retransmission
15PSNR vs. Packet Loss Rate (default ErC)
(Foreman CIF)
16PSNR vs. Packet Loss Rate (Adaptive ErC
default ErC)
(Foreman CIF)
17Visual Quality
Thumbnail-aided retran. default ErC (26.7 dB)
Thumbnail-aided retran. adaptive ErC (27.0 dB)
Experiment settings 20 packet loss, 100
bandwidth
18Conclusions
- Thumbnail-aided distortion-aware retransmission
can achieve gains of 0.5 1.5 dB over
distortion-unaware heuristic methods - Thumbnail-aided adaptive error concealment can
achieve additional gains of 0.5 1 dB under
severe distortion conditions - Key ingredient leading to gain is content-level
error detection and correction
19Acknowledgement
- I would like to thank Prof. Girod, Yao-Chung Lin,
Xiaoqing Zhu, David Varodayan and Pierpaolo
Baccichet for extremely helpful discussions.
20Basic Settings
GOP
Video Sequence
I
B
P
B
P
B
Picture
Packet
0101
Thumbnail
Slice
- Packetization
- one slice per packet
- Thumbnail generation
- I and P pictures only
- For example, 32x32 block maps to 1 pixel of 4-bit
in thumbnail
21Analysis of Two Projections
- Analysis model
- Channel distortion non-zero mean (to
characterize local distortion) and s.d. ,
i.i.d. - Want to estimate the noise power
- Mean-based estimator
- Random-projection estimator
-
- where with equal prob.
22Analysis of Two Projections (Contd)
- Mean-based estimator
- Mean
i.e. biased, always
underestimate - Variance as
- Random-projection estimator
- Mean , i.e. unbiased
- Variance as
i.e. large variance if noise is
not zero-mean
23PSNR vs. Packet Loss Rate (w/o Error
Concealment)