1. Presentation of the paper: Multiple Description Coding: Compression Meets the Network by Vivek K Goyal Angelos Chatzipapas Alper Kivrak Maria Michalopoulou

2. Problem: –The web browser repeatedly gets stuck with partially loaded web pages Solution: –More antennas on laptops => higher throughput and better reliability –More cellular base stations imply smaller cells => fewer conflicts with other users –Wireless service provider could provide higher capacities in the wired connections to its based stations => fewer packets lost due to buffer overflows Introduction (1) _____________________________________________________________________________________________ Seminar in Communications Engineering

3. Focus of this article: Compressed representations of the pictures The representation does not affect how many bits get from the web server to your laptop, but it determines the usefulness of the bits that arrive. Introduction (2) _____________________________________________________________________________________________ Seminar in Communications Engineering

4. Current systems typically generate content with a progressive coder and deliver it with TCP In a progressive transmission the quality improves steadily as the number of consecutive packets received, starting from the first, increases The order of the packets is critical Progressive transmission works well when the packets are sent and received in order without loss Where does a conventional system go wrong?_____________________________________________ Seminar in Communications Engineering

5. Background: –MD coding was invented at Bell Laboratories in connection with communicating speech over the telephone network Telephone system = reliability but outages of transmission links are inevitable –1970: the primary mechanism for providing uninterrupted service despite link outages was to divert calls to standby transmission links => increases cost and not all the installed links are used in normal operation MD Model - History (1) _____________________________________________________________________________________________ Seminar in Communications Engineering

6. –To improve reliability without standby links => split the information from a single call and send on two separate links or paths in normal operation both halves of the transmitted data would be combined for the usual voice quality an outage of one link would allow for communication at reduced quality MD Model - History (2) _____________________________________________________________________________________________ Seminar in Communications Engineering

7. In telephony, speech signal is bandlimited to 3.2 kHz and sampled at 8 kHz Jayant: sampling at 12 kHz so that subsampling by a factor of two results in only slight aliasing. The odd- and even-numbered samples are compressed separately with differential pulse code modulation and sent on two separate channels. Decoding from both channels requires a DPCM decoder for each channel and the interleaving of the samples, resulting in a signal with 12 kHz sampling and some amount of DPCM quantization noise. To decode from a single channel, adaptive linear interpolation is used. MD Model - History (3) _____________________________________________________________________________________________ Seminar in Communications Engineering

8. During “Channel Splitting”, following question comes to mind, “If an information source is described with diferrent descriptions, what are the limitations on qualities of these descriptions taken separately and jointly?” This is known as MD problem MD Problem (1) _____________________________________________________________________________________________ Seminar in Communications Engineering

9. MD situation with two descriptions; Scenario for 2 channels and 3 receivers General case has M channels and 2^(M-1) receivers MD Problem (2) _____________________________________________________________________________________________ Seminar in Communications Engineering

10. MD coding is difficult because of conflicting requirement; If we have good description at rate R1 and R2, no need to spend R1+R2 total bits Just use better one Also, good compressed R1+R2 is difficult to split into two Making descriptions individually good is the fundamental trade of MD coding MD Problem (3) _____________________________________________________________________________________________ Seminar in Communications Engineering

11. So, MD Model leads to several problems in rate distortion theory and practical compression Problem to determine corresponding values (R1,R2,D0,D1,D2) Aim is to code with no more total bits than would be necessary to communicate the source in one description This is the original motivation of channel splitting MD Problem (4) _____________________________________________________________________________________________ Seminar in Communications Engineering

12. Successive Refinement Coding Special case of MD coding Decoder 2 is removed Information on Channel 2 is not useful without Channel1 SR Coding (1) _____________________________________________________________________________________________ Seminar in Communications Engineering

13. Important data should be sent on Channel1 Additional data on channel1 improve reconstruction quality For example: for images, low pass version can be sent on channel1 and additional details can be sent on channel2 This situation occurs in packet communication when multicasting to users with different available bandwidths SR Coding (2) _____________________________________________________________________________________________ Seminar in Communications Engineering

14. MD coding is unlikely to be useful under absence of following conditions –One or more users sometimes fail to receive one or more descriptions. –Various distortions are acceptable and distinguishable. Distortion measure must reflect utility. Central distortions lower than side distortions should create some value. MD Coding Conditions _____________________________________________________________________________________________ Seminar in Communications Engineering

15. Where to Use MDs (1) _____________________________________________________________________________________________ Seminar in Communications Engineering Distributed Storage Systems Distributed Storage Systems  Distributed Storage systems are rapidly evolving - Internet is best example; we access every day Web pages whose display requires data from dozens of machines around the world.  The concept of Distributed Storage matches the MD framework incredibly well.  Example Scenario: Downloading images from an Internet image library. -Case 1: identical data is stored at the servers. Potential Congestion very high delays or even total failure -Case 2: low resolution versions of the image at multiple sites Potential Congestion users have to satisfy themselves with a lower resolution version of the image reconstructed using the sites that become available within a reasonable delay.

16. Where to Use MDs (2) _____________________________________________________________________________________________ Seminar in Communications Engineering In Packet Networks In Packet Networks When and Why?  Retransmission mechanisms are not always possible -in wireless environments the links are not always bidirectional - in real time communications the added delay is not acceptable (e.g. video streaming, voice transmission etc)  FEC performance is not sufficient if the number of packets is not large enough The implementation of MD in packet transmission is quite straightforward and thus, this possibility is mentioned in many recently published works related to MD techniques

17. Where to Use MDs (3) _____________________________________________________________________________________________ Seminar in Communications Engineering unit interval 1 … 2 3 4 … TIME L Frequency Hopping Wireless Systems Frequency Hopping Wireless Systems  Frequency hopping provides protection against a long-term interference or long-term fading at the location of the receiver. Error characteristics are constant within unit interval The L carrier frequencies can be considered separate channels for and MD source code. FREQ channel

18. Where to Use MDs (4) _____________________________________________________________________________________________ Seminar in Communications Engineering Hybrid Digital Broadcast Hybrid Digital Broadcast  At the edge of a radio station’s range the interference from a station with neighboring frequency may be significant. 200MHz interference  Adding digital information to radio broadcasts for eliminating the interference at the edges of the band  But, some users have interference at the upper edge and others have it at the lower edge.  Thus, these edges make for good MD channels; some users will get only one or the other channel and other users will get both.

19. For what types of Data? The data must be useful at various quality levels The data must be useful at various quality levels _____________________________________________________________________________________________ Seminar in Communications Engineering -Generally, it can be applied to most data that are lossily compressed audio images video etc -It would not be useful for: text numerical data etc

20. Summarizing… _____________________________________________________________________________________________ Seminar in Communications Engineering MD Coding in a few words: MD Coding in a few words: - - fragments a single media stream into n independent sub streams - in order to decode the media stream, any description can be used, however, the quality improves with the number of descriptions received in parallel When and Why to use MD - in many cases, it is better to have lower quality data than no data ! - So, when the transport mechanism is not perfect, compensate this with appropriate coding !

21. That’s all! That’s all! Thank you for your attention!