1 / 47

Prioritized Distributed Video Delivery With Randomized Network Coding

Prioritized Distributed Video Delivery With Randomized Network Coding. IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 13, NO. 4, AUGUST 2011 Nikolaos Thomos Jacob Chakareski Pascal Frossard. Outline. Introduction Network Coding and UEP Video Streaming Distortion Analysis Optimal Rate Allocation

zizi
Download Presentation

Prioritized Distributed Video Delivery With Randomized Network Coding

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Prioritized Distributed Video Delivery With Randomized Network Coding IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 13, NO. 4, AUGUST 2011 NikolaosThomos Jacob Chakareski Pascal Frossard

  2. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  3. Introduction • Because the advances in broadband technologies, video compression, and ever-increasing amount of multimedia content, the development of new delivery architectures has been accelerated. • Peer-to-peer systems have experienced a fast development and emerged as one of the most popular online media delivery

  4. Overlay Networks • In such media streaming systems, networks are often organized in overlay structures that provide better control of the delivery process.

  5. Overlay Networks • Network coding can be offered at intermediate network nodes in overlay networks. • increase source diversity. • improve delivery performance. • increase the goodput of transmission.

  6. Heterogeneous Peers • The growing heterogeneity of Internet access links’ characteristics • packet loss • bandwidth • It’s has created an important need for scalable delivery mechanisms.

  7. Multimedia Data • Multimedia data is typically characterized by a variable importance of the data units in terms of their contribution to the overall reconstructed quality. • The delivery should be organized such that the peers are served the data efficiently according to their capacity.

  8. Problem • This paper addresses the problem of prioritized media streaming in overlay networks, where network coding operations are designed to cope with media packets of different importance.

  9. Proposed Scheme • This paper proposes an efficient streaming scheme that allows for multiple levels of quality-of-service in order to accommodate for the network heterogeneity.

  10. Randomized Network Coding • This paper builds on the results of randomized network coding (RNC) [9] for the construction of a distributed streaming solution that improves the robustness to erasures without the need for centralized control, as described in [10]. [9] T. Ho, R. Koetter, M. Medard, D. R. Karger, and M. Effros, “The benefits of coding over routing in a randomized setting,” in Proc. IEEE Int. Symp. Information Theory, Kanagawa, Japan, Jul. 2003. [10] P. A. Chou, Y. Wu, and K. Jain, “Practical network coding,” in Proc. 41st Allerton Conf. Communication Control and Computing, Monticell, IL, Oct. 2003.

  11. Contributions • This paper proposes • a new distributed delivery algorithm where the coding decisions are adapted to prioritized video delivery for receivers. • a receiver-driven network coding strategy where the receiving peers request packets from classes with varying importance.

  12. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  13. Network Coding • The nodes in the network perform linear combinations of the received packets and transmit the coded packets to the destination nodes. • The receivers can then recover the original data by receiving and subsequently decoding a sufficient number of linearly combined packets.

  14. Network Coding • Distributed algorithms are considered where each node independently chooses its coding strategy based on a local network view. • RNC [9], the coding coefficients are selected randomly by network coding nodes. • Without any need for a central coordination. • Without comprehensive knowledge of the network topology.

  15. Network Coding • The network coding operations can be written as follows. If a node u generates Mpackets by RNC, then the mthnetwork coded packet cmis of the form

  16. Decoding Latency • The packet stream is split into multiple generations and coding operations are restricted to packets within the same generation. • Typically, a generation can correspond to a GOP. • The generations are characterized with playback deadline information, the network nodes only transmit useful packets with decoding deadlines have not passed.

  17. Proposed Prioritized RNC • The packets are organized into C classes depending on their importance. • The class cis defined as the set of packets that are linear random combinations of packets from the cmost important classes. Base layer Enhancement layer

  18. Proposed Prioritized RNC • Their objective is to design a novel network coding algorithm in overlay networks that is able to • deal with packets of different importance. • increases the likelihood of delivery for higher priority packets. • The mixing operations should not be uniform across all packets arriving at a node, but instead packets with higher importance should be involved in more coding operations.

  19. Proposed Prioritized RNC • The receiver-driven policy provides a simple way to adapt to the capabilities of the peers. • The children peers send requests to their parents. • specify the expected number of packets from each importance class. • The coding operations are driven by the children nodes that determine the optimal amount of coding allocated to each importance class of the data to which the network nodes subscribe.

  20. Packet Delivery Protocol • The children nodes compute the optimal coding strategy that should be implemented at their parent nodes, based on the available network bandwidth, the expected loss probability, and the importance of packets in each class.

  21. Packet Delivery Protocol • The parent nodes in turn randomly combine their packets according to the computed coding strategies and forward the corresponding coded packets to their children.

  22. Packet Delivery Protocol • A child node finally inspects the incoming packets to determine whether they are innovative. Non-innovative packets are removed from the node’s buffer. • Based on the state of its buffer and the local network status, the child node then computes again the optimal coding strategy and sends it to its parent nodes. • This procedure is repeated periodically.

  23. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  24. Distortion Analysis • The distortion is dependent on the number of classes that can be decoded. • The number of native video packets in the first cclasses is written as • The probability of decoding a class depends on the number of received coded packets. • A client is able to decode the cthclass as soon as it receives innovative βcnetwork coded packets.

  25. Distortion Analysis • The expected number of packets of class c sent by the node u is given as

  26. Distortion Analysis • This paper actually consider the maximum value between the incoming and outgoing bandwidths at a peer node as the capacity constraint in our network coding algorithm.

  27. Distortion Analysis • The packet loss probability is equal to

  28. Distortion Analysis • The distortion experienced at node is simply written as Dmax– D(u), where Dmaxrepresents a constant maximal distortion when no video class can be decoded. Back

  29. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  30. Rate Allocation Problem • The node is interested in determining the number of packets it should request from its parents for each packet class. • The optimal class distribution ω* is computed by each client node such that it minimizes the expected distortion, as contributed by the received packets.

  31. Rate Allocation Problem • The optimal rate allocation problem can be written as follows: • The peer is interested in selecting the vector of coefficients such that the network coding strategy employed by its parent nodes maximizes the received video quality.

  32. Optimization Algorithm • Every client peer has to solve the RAP problem independently based on local network information. • Since the search space is huge, exhaustive algorithms are too complex and cannot be implemented realistically.

  33. Log-concave Function • The objective function is a log-concave function, which leads to a simple iterative solution of the problem in each peer node.

  34. Optimization Algorithm Supplement

  35. Optimization Algorithm • Each client peer runs the above optimization problem periodically and requests the optimal distribution ω*from its parent nodes. • The parents then implements network coding operations in order to match the requested distribution. • Note that it might happen in practice that parents are not able to transmit the requested packets. • the parent node distributes ω* uniformly to the classes it can transmit.

  36. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  37. Simulation Results • Video source : Foreman sequence • encode with the scalable extension (SVC) of the latest video compression standard H.264 [13]. • encode in CIF image size and three quality layers. • The size of the GOP is set to 30 frames. • The frame rate is 30 fps. • The packets of 1500 bytes are augmented by the TCP/IP and the network coding headers. • Each evaluation point in our analysis is the average performance computed over 100 network topologies with similar statistical properties.

  38. Full Search V.S Optimal Rate Allocation Search

  39. Decoding Performance [25] D. Sejdinovic, D. Vukobratovic, A. Doufexi, V. S. , and R. Piechocki, “Expanding window fountain codes for unequal error protection,” in Proc. 41th Annual Asilomar 2007 Conf. Signals, Systems and Computers, Pacific Grove, CA, Nov. 2007.

  40. Decoding Performance [25] D. Sejdinovic, D. Vukobratovic, A. Doufexi, V. S. , and R. Piechocki, “Expanding window fountain codes for unequal error protection,” in Proc. 41th Annual Asilomar 2007 Conf. Signals, Systems and Computers, Pacific Grove, CA, Nov. 2007.

  41. Performance Under Timing Constraints

  42. The Influence of Buffer Size

  43. Outline • Introduction • Network Coding and UEP Video Streaming • Distortion Analysis • Optimal Rate Allocation • Simulation Results • Conclusion

  44. Conclusion • This paper has proposed a novel receiver-driven RNC technique with built-in UEP properties. • Also considers the unequal importance of the various packet classes and implements different random network coding protection levels. • The UEP properties are achieved simply by choosing the proper rate allocation among the different classes.

  45. Optimization Algorithm • The optimization algorithm starts from a pivotal packet distribution ωover the priority classes that is then refined iteratively. • The initial distribution depends on the number of classes that a node can optimally decode given the overall number of received packets. • In every step of the algorithm, we examine the neighbors of the distribution ω vector obtained from the previous iterations.

  46. Supplement Optimization Algorithm • For example, if the video is encoded with packets in three classes arranged in order of importance, we can write the lthcandidate distribution vector of the iterative search algorithm at step tas Back

  47. Supplement Optimization Algorithm • The algorithm checks the expected reduction in distortion for each of the neighbor rate allocations. If one of them results into a larger reduction in distortion than the starting rate allocation rt, l(u), then the neighbor allocation is included in the list of candidate solutions. • This procedure is repeated for all new candidate solutions: the neighbor allocations are tested, as long as the unit rate transfer between the priority classes decreases the overall distortion. When there are no further beneficial packet exchanges, the algorithm stops and the best candidate solution is retained.

More Related