1. Video Quality Research @ IBCN

2. 2  Real-time Video Quality Monitoring/Testing  Monitor Probe  Virtual Wall  Video Quality Metrics  Full Length Movie Quality Assessment  Visual Quality Impairment Detector  VQEG involvement  Video Streaming  xStreamer

3. 3 Real-time Video Quality Monitoring/Testing  Monitor probe  Real-time monitoring of H.264/AVC streams at several demarcation points  Gather both network statistics and video statistics  Packet loss, delay, jitter  Macroblock & motion vector information  N. Vercammen, N. Staelens, B. Vermeulen and P. Demeester, “Distributed Video Quality Monitoring”, to appear in Proceedings of 2nd IEEE International Workshop on Internet and Distributed Computing Systems (IDCS'09), December 10-12, Korea

4. Monitor probe 4

5. 5 Real-time Video Quality Monitoring/Testing  Virtual Wall & Video testbed  Video testbed  Automate video quality tests  Virtual wall  Enables creation of multiple video testbeds  N. Vercammen, N. Staelens, B. Vermeulen and P. Demeester, “Extensive video quality evaluation: A scalable video testing platform”, Proceedings of 1st IEEE International Workshop on Internet and Distributed Computing Systems

6. Scalable Video Testing Platform 6

7. 7 Video Quality Metrics  Full Length Movie Quality Assessment  Existing video quality assessment methodologies:  Evaluate short video sequences (~ 15s)  Users are actively evaluating visual quality  Overall test duration limited to 30 minutes  Watching television  At home, living room, with family (social viewing)  Longer content: movies, television programs  Lean backward TV viewing experience  => how is quality perceived while watching full length movies, when users are not focused on quality evaluation

8. 8 Full Length Movie Quality Assessment  Error Visibility  Frame freeze in movie frame freeze in short sequences: 42% 91%  Blockiness in movie and short sequences: both 98%  Error Annoyance  Standard test: freezes rated higher quality than blockiness  Movie: frame freezes are more annoying  Conclusion  Focus is important  Flow experience

9. 9 Full Length Movie Quality Assessment  N. Staelens, B. Vermeulen, S. Moens, J.-F. Macq, P. Lambert, R. Van de Walle and P. Demeester, “Assessing the influence of packet loss and frame freezes on the perceptual quality of full length movies”, Proceedings of Fourth International Workshop on Video Processing and Quality Metrics for Consumer Electronics (VPQM-09)  N. Staelens, S. Moens, W. Van den Broeck, I. Mariën, B. Vermeulen, P. Lambert, R. Van de Walle and P. Demeester, “Assessing the perceptual influence of H.264/SVC signal-to-noise ratio and temporal scalability on full length movies”, Proceedings of First International Workshop on Quality of Multimedia Experience (QoMEX 2009)

10. 10 Video Quality Metrics  Visual Quality Impairment Detector  Current research focus  Real-time visual impairment detection  Network level => detect losses  Video level => determine severity and visibility  Target No-Reference metric or Reduced-Reference  Will be implemented in the monitor probe

11. 11 VQEG involvement  Ghent University – IBBT involved in VQEG  Meeting hosted, September 22 – 26, 2008  Contributes to  MM testplan (already finalized)  ITU-T J.246 & ITU-T J.247  Hybrid testplan:  co-editors  HDTV testplan  Independent Lab Group (ILG)  Joint Effort Group (JEG)  Toolchain for creating impaired sequences  H.264/AVC parser, based on JM reference software

12. xStreamer Modular Multimedia Streaming

13. xStreamer  In-house developed modular multimedia streamer  Alexis Rombaut (alexis.rombaut@intec.ugent.be)alexis.rombaut@intec.ugent.be  Written in C++  Uses libraries:  libavformat/libavcodec (parsing/encoding/decoding)  live555 (RTSP)  jrtplib (RTP)  Released under General Public License (GPL)  Freely available at http://xstreamer.atlantis.ugent.be/http://xstreamer.atlantis.ugent.be/ 13

14. 14 Modular Multimedia Streaming Inspired by Click Modular Router & DirectShow Offers different components  Performs basic functions  Readers, packetizers, multiplexers, schedulers, transmitters, receivers, writers, classifiers, analyzers Streamer is directed graph of components

15. 15 Modular Multimedia Streaming Supports audio and video Using RTP packetization:  MPEG-1/2/4 Video & Audio Using MPEG-2 Transport Streams:  MPEG-1/2/4 Video & Audio  H.264 AVC/SVC

16. 16 Modular Multimedia Streaming Multitude of supported protocols RTP/UDP RTSP/RTP/UDP UDP TCP

17. What can xStreamer do? Advanced streaming server  Own MPEG-2 TS multiplexer  SVC streaming  Differentiated streaming using classifiers Proxy/client  Proxy: convert differentiated stream into a single stream  Client: save captured stream to file Video tool  No ‘real’ streaming involved  Simulate packet loss  Collect tracefiles during streaming 17

18. Create xStreamer configuration Configuration saved in XML-based file  Describes directed graph of components and connections between components Graphical User Interface  Visualize directed graph  Drag components and draw connections  Configure components 18

19. Example: Differentiated SVC streaming 19 Read raw H.264 video stream Packetize frames into packets as defined in RFC3984 Avoid bursts by smoothing packets over time Classify NAL units depending on SVC layer Stream different layers over different connections

20. Example: Proxy/Client 20 ProxyClient

21. xStreamer as video tool Offline simulator  No ‘real’ streaming  Simulate packet loss using Classifier component  Random, Gilbert-Elliott  Write resulting packet stream back to file Tracefile generation  Packetizer : video trace  Transmitter : sender trace  Receiver : receiver trace  Classifier : sender & receiver trace 21

22. Publication  ‘xStreamer: Modular Multimedia Streamer’ accepted for publication on ACM Multimedia 2009 - Open Source Software Competition, Beijing, China October 19-24, 2009

23. Distributed transcoding with xStreamer

24. Current: Architecture Transcoder 1 Transcoder 1 Proxy Transcoder … Transcoder … Transcoder N Transcoder N Server Input File Input File Output File Output File

25. Current: Server Differentiator ffmpeg- reader ffmpeg- reader RTP transmitter 1 RTP transmitter 1 RTP transmitter … RTP transmitter … RTP transmitter N RTP transmitter N Send each new GOP to the next transcoder = RR (round robin) distribution note: the schemes omit some components for clarity

26. Current: Transcoder xstreamer unpacketizer xstreamer unpacketizer RTP receiver RTP receiver transcoder xstreamer packetizer xstreamer packetizer RTP transmitter RTP transmitter

27. Current: Proxy ffmpeg- writer ffmpeg- writer multiplexer RTP receiver 1 RTP receiver 1 RTP receiver … RTP receiver … RTP receiver N RTP receiver N

28. First experimental results  The following slides show the results from the first experiments on the virtual wall.  Experiment parameters:  All sequences are 90 minutes long, encoded with H.264 at 25 frames per second.  Future work will experiment with different bit rates for each resolution. SequenceResolutionBitrate (kbps) QCIF176x144100 CIF352x288400 4CIF704x5761600 720p1280x7203600 1080p1920x10808100

29. Scalability  The figure shows the how much faster than real-time (factor) we can transcode in function of the number of nodes.  For example, using 20 nodes the system can transcode from 1080p to QCIF 100 times faster than real-time, transcoding 90 minutes of video in less than 1 minute.  Some curves flatten because the server cannot stream more than 1 Gbps to feed the transcoding notes, future work will alleviate this by using several network interfaces.

30. Codec comparison  Provides the same information as the previous figure but organized in function of the source resolution.  As target resolutions become larger the influence of the source resolution decreases (encoding, determined by target resolution, consumes more resources than decoding, determined by source resolution).  For example transcoding 1080p to 4CIF is not much slower than transcoding 720p to 4CIF (blue bars).

31. Transcoding duration  The figure shows the considerable processing power by transcoding 90 minutes of video in mere 10 seconds for the smallest resolutions.  For the highest resolutions, the system can transcode 90 minutes of video from 1080p to 720p in less than 4 minutes.  The times to transcode 1080p to QCIF and 1080p to CIF are same because the server could not feed the former adequately.

32. Conclusion  The first experiments show promising results by transcoding between 25 and 500 faster than real-time depending on the resolutions using 32 nodes.  Some combinations did not fully scale up to 32 nodes because the server bit rate would exceed 1 Gbps. However, future work, using multiple interfaces or multiple servers will avoid this.  Future experiments will increase the number of nodes up to 100.