2. SHEAU-RU TONG Management Information System Dept., National Pingtung University of Science and Technology, Taiwan (R.O.C.) srtong@mail.npust.edu.tw YUAN-TSE YU Dept. of Software Engineering National Kaohsiung Normal University, Taiwan (R.O.C.) yuyt@nknucc.nknu.edu.tw 1

3.  Background  Replicate Multiple Descriptor (RMD) Coding  Simulation and Performance Evaluation  Conclusions 2

4.  Multi-path fading and fast interference in wireless networks causes heavy packet loss burst and poor video broadcast quality.  Low-layer forward error coding (FEC) is suitable for fixing short-term loss burst, but not for long- term loss burst. 3

5.  A HSDPA traffic trace (with link-layer FEC) of a moving vehicle where multiple-packet-loss events frequently occur. 4

6.  Multiple descriptor coding (MDC): Split a video stream into several sub-streams (descriptors) Encode each sub-stream independently and transmit them over different data paths. Exploiting path diversity.  Protecting key-frame is more efficient! 5 Application-layer error resilience is desirable!

7.  Basic idea: logically applying MDC in the application layer with two new features. Replicate key-frames Interleave multiple descriptors over an IP multicast channel.  Advantages: Distributing long loss burst over multiple descriptors to mitigate the damage impact. Exploiting time diversity of key frame transmission. 6

8. The system architecture of RMD 7

9.  Frame replication/insertion b: the number of replicas. s: the number of the frame slots shifted. 8

10.  Analysis of RMD Frame slot time (I sub-stream ) where k is number of sub-streams, r video data rate and p GOP size. Delay (D) and buffer space (B) where S I and S O are the maximum frame size of I frames and the other frames, respectively. 9

11. Key frame error probability (P kf_err ) where  is the packet slot error rate, g is the number of packets for an I frame. 10 P­ kf_err with respect to various b’s when k=4 and g=4.

12.  Simulation configurations 400 frames MPEG-4 with a QCIF format. 30 fps with GOP pattern of IPPPPPP. Packet size of 1024 bytes  Comparing RMD with Single Descriptor (SD) Multiple Descriptors (MD) SD/FEC(255, 159) MD/FEC(255, 159) 11

13. 12  When the packet loss rate increases beyond 15%, RMD starts to outperform SD/FEC or MD/FEC (PSNR>21dB).

14. 13  The instances of PSNR with respect to various schemes when packet loss rate  = 20%. RMD protects the key frame better!

15. 14 Average PSNR of RMD with respect to various s’es when b=3. Average PSNR of RMD with respect to various s’es when b=2.  Impact of b and s in RMD The PSNR curve is improved (leveraged) as s and b increases. The improvement is magnified under a heavy packet loss condition (15%-35%)

16.  Redundancy cost (ε =20%) RMD (b=3) has a redundancy ratio slightly higher than that of MD/FEC, but can gain a PSNR about 5 dB. 15

17.  RMD is an ideal application layer approach for combating excessive burst errors and protecting key frames.  RMD (b=3) has a redundancy ratio slightly higher than that of MD/FEC, but can gain a PSNR about 5 dB subject to ε =20%.  How to co-work RMD with other adaptive FEC to offer a full spectrum of protection against various burst errors is worthy of further study. 16