1. Error Resilience in a Generic Compressed Video Stream Transmitted over a Wireless Channel Muhammad Bilal 2005-06-0020

2. Channel Noise  Markov Process Model  AWGN Model  Transfer Function Model  Hata Model, Akumura Model

3. Error Correction Methods Redundancy Redundancy  Header information  Motion Vectors  DC coefficient Source Coding Source Coding  Reed Solomon  Hamming Code Channel Equalization Channel Equalization  Channel coding  Channel response

4. Error Localization Reversible Variable Length Codes Reversible Variable Length Codes Fixed Synchronization Markers Fixed Synchronization Markers

5. Error Resilience Methods Data Partitioning Data Partitioning SNR scalability SNR scalability Prediction Prediction –Intra coded frame  Copy previous block  DC coefficient prediction –Inter coded frame  Motion Vector Data based prediction

6. Framework A generic video compressor A generic video compressor –MATLAB implementation –‘MPEG-2 like’ bit stream –Platform for video coding analysis  Compression efficiency  Motion Estimation (offset distortion)  Data Partitioning  Etc –Demonstration of good quality & highly quantized videos

7. Error Introduction Methods Arbitrary bursts of error in bit stream Arbitrary bursts of error in bit stream –Header loss –RVLC synchronization problem –Need to deal with resynchronization

8. Error Introduction Methods (contd.) Intelligent error introduction (Macroblock level) Intelligent error introduction (Macroblock level) –Assume bit stream remains synchronized –Error in coefficients/motion vector data –SNR degradation –Demonstration  Error propagation due to motion compensation  Need for ‘I’ frame GDR (Gradual Data Refresh)

9. Quality Measures Subjective evaluation Subjective evaluation SNR SNR –Deceiving results for some sequences

10. Analysis Effect of various error concealment methods Effect of various error concealment methods –I Frames  DC Coefficients saved –‘D’ frame  DC Coefficients not saved –Copy previous frame block  Error propagation due to motion compensation

11. Analysis (contd.) ‘P’ Frames ‘P’ Frames –Dependent on ‘I’ frame (error propagation) –Dependent on content  High motion content (Foreman)  Head & Shoulder (News)  Camera panning (Coastguard) –Motion Vector Data + DCT  DCT data useless without MV  MV data useful without DCT data  demonstration

12. Analysis (contd.) ‘P’ Frames ‘P’ Frames –Absence of DCT data  Copy motion compensated block  Previous frame non MC block degrades video for high motion content

13. Analysis (contd.) ‘P’ Frames ‘P’ Frames –Dependency on ‘I’ frames –Perfect ‘I’ frame decoding  DCT data destroyed  MV data available  Demonstration –Seamless video (news) –Acceptable video for many purposes (coastguard, foreman)

14. Analysis (contd.) Critical Data Critical Data –I Frame  DC coefficients –P Frame  I frame  MV data  Motion Estimation algorithm –Attempt to find the ‘actual’ motion vector

15. SNR vs BER

19. Conclusion Error Resilience Error Resilience –Error localization  Parity  Hamming Codes –Redundancy  DC coefficients + MV data  I frame perfect decoding (BW demanding) –Compensate with more number of P frames in GOP

20. Conclusion (contd.) Data Partitioning Data Partitioning –Critical data positioned close to resynchronization marker  DC coefficients in ‘I’ frames  MV data in ‘P’ frames

21. Conclusion (contd.) Further compression! Further compression! –Randomly introduce ‘not coded’ blocks  Depend on decoder error concealment scheme  Infrequent ‘not coded’ blocks will result in seamless video decoding

22. Q&A