Title: Low Power Compression for Mobile Video Recording Using a Combination of H'264 and WynerZiv Coding
1Low Power Compression for Mobile Video Recording
Using a Combination of H.264 and Wyner-Ziv Coding
2Outline
- Introduction- MORE DESCRITPITON
- Combination of H264 and Wyner-Ziv
- Simulation Result
- Future Work
3Application Scenario
4Introduction
- Conventional Codec
- High complexity encoding
- Low complexity decoding
- Wyner-Ziv Codec
- Low complexity encoding
- High complexity decoding
- Combined Scheme
- Base Layer
- Coding low resolution conventionally
- Enhancement Layer
- Coding high resolution in Wyner-Ziv scheme
5Goal
- Lower overall complexity than conventional codec
- Power saving in encoding
- Real-Time demonstration
- Realized in TriMedia Platform
- Show the importance of temporal interpolation
- Benefit from the Natural Motion in TV technique
6Outline
- Introduction
- Combination of H264 and Wyner-Ziv
- Simulation Results
- Future Works
7Coding Scheme
Bernd Girod, Low-Power Compression for Mobile
Video Recording Using a Combination of H.264 and
Wyner-Ziv Coding, June 2005
8More des.
- Rate control of WZ with BL Modes MVs
- Bit allocation between BL and EL
- Side information
- Gain of using interpolation error
- Rate distortion Performance
- Complexity
9 change title
- Set up simulation environment
- Design interfaces
- Integrate H.264 as a library
- Investigation channel decoding implementation
- We chose 6-bit quantization PLR (authors year)
- Avoid floating point
- Negligible rate loss
- Represent in byte
- Optimization on TriMedia platform
- Rate-distortion improvement
- Adopting Philips Natural Motion MCI
- Using Decoded Enhancement Layer
10Current Status
11Current Status
- Up/Down Sample ZOH, 6-tap FIR
- H.264 codec TML 90 (Ported as a Library)
- Interpolator
- Temporal Philips Natural Motion
- Spatial ZOH, 6-tap FIR
- Wyner-Ziv
- H264 Intra quantization, 4x4 Transform
- LDPC channel coding
12Wyner-Ziv Codec for Video Signal
13Use of Enhancement Layer
14Use of Enhancement Layer
Prediction only from base layer
Prediction with Enhancement Layer
15LDPC Decoding Consideration
- For Trimedia Platform
- 5 issue slots
- 2 load/store units
- 16k Cache
- SIMD instructions
- Prefer
- Fix points operation
- lt8 bits for message
- Use 6-bit quantization table lookup (4kbytes)
2
- Or
- 1 load/store units
- 128K Cache
16Required Throughput
17Throughputs of different Implementations
Src Architecture and Implementations of
Low-Density Parity Check Decoding Algorithms,
Midwest Symposium on Circuit and System, 2002 3
93kbps
46Mbps
1.1Mbps
18Video Codec Speed
- Coding 15 frames of Foreman CIF
- B24, E12 (5 bit planes)
- LDPC block size 6336 for Luma, 396 for Chroma
- Without optimization on PC
Note H264 coding QCIF
19Outline
- Introduction
- Combination of H264 and Wyner-Ziv
- Simulation Results
- Future Works
20Simulation Results
- Rate-Distortion Performance
- Different Prediction Modes
- (Symbols too small)
21Compare with/ without FIR
foreman CIF 31 frms PSNR of Luma Rate
Entropy Quant Set 4,12, 18, 24, 28
22Setup
- Video Source CIF, 31 frames
- Distortion PSNR of Luminance
- Rate Entropy
- Up-sample FIR
- Temporal Interpolation SNM
- Quantization Parameter
- E,B 4,12,18,24,28
23Prediction Modes
24Notation
- P picture
- the picture has corresponding lower resolution
picture - B picture
- the picture has no corresponding lower resolution
picture - Pred_XY
- X the prediction mode of P picture
- Y the prediction mode of B picture
- 0 from base layer, 1 from enhancement layer
- Pred_mv
- Using Pred_11, but no residue coded at base
layer, independent from base layer except using
mv and modes. - Pred_1p
- P means perfect for B picture
25Side Information
(a) Up sampled base layer
Motion Compensated (b)
(c) MCI of EL
MCI of BL (d)
26Coded Rate vs. Frame
27(No Transcript)
28(No Transcript)
29(No Transcript)
30Summary
- 3 dB gain with FIR spatial filter
- 1 dB gain from NM-MCI of enhancement layer
- 0.5 dB gain from MC at low bit rate base layer
- 0.5 2 db by removing base layer redundancy
- 4 dB Rate-Distortion gap from H264
- 1dB gap from H264 in ideal case at high bit rate
31Outline
- Introduction
- Combination of H264 and Wyner-Ziv
- Simulation Results
- Future Work
32Future Work
- Rate Distortion Optimization
- Discover the correlation between BL mode, mv, and
rate - Base Layer H264 CBP, skip modes
- Adaptive spatial up-sample filter
- Rate Control (Estimation) at Encoder
33Rate Distortion Optimization
- Using predicted signal (from side info b1bn-1)
- Choose not to code if -?R/?D gt ?
34Rate Distortion Optimization
35Adaptive Up-Sample Filter
- Learn the partitions from previous pairs
- Adapt the coefficients of up-sample filter
- Discover the correlation between base and
enhancement layer at decoder - Make use of enhancement layer at spatial direction
36Conclusion Discussion
- Set up Initial Simulation Environment
- Optimize Rate-Distortion Performance
- Benefit from Natural Motion
- Lower the complexity
- Benefit from TriMedia Platform
37Acknowledgement
- Thanks Prof. Girod for this great idea
- Thanks Erwin, Torsten, Mark, Maurice, and Ralf
for helpful discussions and suggestions.
38Reference
- 1 Bernd Girod, Research Proposal Low Power
Compression for Mobile Video Recording Using a
Combination of H.264 and Wyner-Ziv Coding, June
2005 - 2 L. Ping and W.K. Leung, Decoding Low Density
Parity Check Codes with Finit Quantization Bits,
IEEE Communication Letters, Vol. 4 No. 2, Feb
2000 - 3 E. Yeo, B. Nikolic, and V. Anatharam,
Architectures and Implementations of Low-Density
Parity Check Decoding Algorithm, Midwest
Symposium on Circuit and System, 2002 - 4 W. E. Ryan, "An Introduction to LDPC Codes,"
in CRC Handbook for Coding and Signal Processing
for Recording Systems (B. Vasic, ed.) CRC Press,
2004
39Q A