1. An Error-Resilient GOP Structure for Robust Video Transmission Tao Fang, Lap-Pui Chau Electrical and Electronic Engineering, Nanyan Techonological University IEEE Transactions on Multimedia, Dec. 2005

2. Outline  Introduction  Proposed GOP Structure  Efficient Reverse-play Functionality of the Proposed GOP Structure  Experimental Results  Conclusion

3. Introduction(1/4)  GOP I, P, B picture  I frame at the beginning of a GOP A basic entry point to facilitate random seeking or channel switching Provide coding robustness to transmission error

4. Introduction(2/4)  Predictive coding techniques Effect of channel errors on the video can be extremely severe  A novel GOP structure for robust video transmission Insert the I frame in the middle of the GOP

5. Introduction(3/4)  Advantage of the proposed GOP structure Improve the robustness of the streaming video without reducing any coding efficiency Provide reverse-play operation with much less requirement on the network bandwidth

6. Introduction(4/4)  MEPG is designed for forward-play operation Predictive processing complicate the reverse- play operation Straightforward but na ï ve approach  Decode the GOP up to the current frame, and then go back to decode from the beginning of the GOP again up to the next frame to be displayed  Require high bandwidth of the network Some works on the implementation of reverse- play  Require much computation or extra decoder complexity

7. Proposed GOP Architecture  Not including B frames in the following analysis for legibility but without loss of generality  I frame will be stored in the reference buffer

8. Optimal Position of I frame in a GOP(1/3)  Built model to describe the effect of error on the decoded video quality PSNR, MSE  Involve complicated parameter estimation Length of error propagation (LEP) Number of impaired macroblocks (MBs)  Totally n+1 frames in a GOP with one I frame and n P frames

9. Optimal Position of I frame in a GOP(2/3)  A lost slice belongs to ： I frame  LEP = n+1 Forward predicted P frames (P k, k ≦ i)  LEP = i-k+1 Backward predicted P frames (P k ’, k ≦ n-i)  LEP = n-i-k+1  Average LEP

10. Optimal Position of I frame in a GOP(3/3)  By solving, we can get the optimal position of I frame in a GOP, i.e.,

11. Analysis of Coding Efficiency, System Delay and Complexity(1/3)  Coding efficiency Temporary dependency depends on the predicting distance and inherent characteristics of the video sequences  Thus, coding efficiency of Proposed GOP is compatible with that of MPEG GOP Proposed GOP with higher quality while consuming fewer bits  LEP is halved by inserting the I frame in the very middle of a GOP

12.  Number of bits used by conventional MPEG and this work of Mother & Daughter QCIF, Foreman QCIF, and Stefan QCIF

13.  Corresponding PSNR value of Mother & Daughter QCIF, Foreman QCIF, and Stefan QCIF

14. Analysis of Coding Efficiency, System Delay and Complexity(2/3)  System delay Sequence, coding and transmitting order is For displaying, maximal delay will be n-i frames Tradeoff between delay and robustness  Optimization problem becomes

15. Analysis of Coding Efficiency, System Delay and Complexity(3/3)  System complexity For conventional coder, the buffer stores one previous frame for reference For proposed GOP, we enlarge the buffers  Store an extra I frame at encoder  Maximal number of frames stored is n-i for orderly display (display )

16. Efficient Reverse-Play Functionality of the Proposed GOP Structure  The proposed approach outperforms the conventional method, where it saves a lot of retransmissions  And the proposed GOP can reduce the required bandwidth significantly

18.  Bandwidth requirement with respect to the reverse play operation for sending (a) Mother & Daughter, (b) Foreman, and (c) Stefan

19. Experimental Results  In the experiment, specify the frames where slice loss happens beforehand Slice losses at frame 7, 19, 33, 46, 59, 72 These frames are the 6 th, 7 th, 10 th, 1 st, 3 rd, 5 th P frames after each I frame For instance, for slice loss in frame 46, 1 st frame after I frame  The error propagate to all the following P frames by conventional MEPG GOP  The error propagate stops at 6 th frame, say frame 51

20. Experimental Result  Average PSNR comparison according to different delay requirements. (a) Mother & Daughter. (b) Foreman. (c) Stefan

21. Experimental Results  Compared with a compliant structure that starts with an I-frame with compliant prediction directions Put more B frames in the beginning of the GOP to improve the robustness Distances between P (or I) frames are not equal Low compression efficiency due to long predicting distance between P (or I) frames Not good to make the distances between the P (or I) frames different too much

22. Experimental Results  The number below each frame represents the LEP if a lost slice belongs to this frame LEP of proposed structure is 34p LEP of compliant structure is 42p

23. Experimental Results

24. Conclusion  A novel GOP structure that inserts an I frame in a GOP to improve the robustness of the video transmission  From simulation results, the proposed GOP outperforms the conventional GOP  But, it inevitably increase the delay and complexity of the system  And the efficient reverse-play operation for the video streaming system of the proposed structure