1 Adaptable applications Towards Balancing Network and Terminal Resources to Improve Video Quality D. Jarnikov

2 Contents Introduction (www) New solutions Subjective evaluation Conclusions and future plans

3 Introduction of in-home network

4 Problem description Video data transmission Video data decoding Network condition good bad Perceived video quality bad good Perceived video quality bad good Resource consumption low high

5 Objective: networked terminal Resource-constrained terminal : CPU Resource-constrained network : bandwidth Wireless network has fluctuations Source Terminal Wireless network

6 Solution Scalable video technique –Choose size of base layer such, that we can almost guarantee the transmission –The enhancement layer is transmitted if there is available bandwidth –The number of layers to be decoded can be chosen for every frame Controller –Choose how much video data (e.g. layers) should be processed. –Optimizes perceived quality when looking at available input data AND available CPU power

7 Past: Conclusions & Future plans The usage of scalable video enables trade-offs between user perceived quality and network and terminal resources A controller can be used to optimized perceived quality with respect to the available CPU power and amount of input data We developed the controller that doesn’t depend on the scalability technique The correctness of controller behavior depends on rightness of parameters Take into account other terminal resources Organize a feedback from the terminal to the source Create MPEG2 to scalable video transcoder

8 Contents Introduction (www) New solutions –System view –Transcoder –Controller –Network sender-receiver –Summary Subjective evaluation Conclusions and future plans

9 System: Present view Source Terminal Wireless network Transcoder Network Sender-Receiver - Terminal - ?

10 Contents Introduction (www) New solutions –System view –Transcoder –Controller –Network sender-receiver –Summary Subjective evaluation Conclusions and future plans

11 Transcoder: General Info Separates video stream on most important information and least important information (refinement) Input: MPEG2 video Output: Scalable MPEG2 video Parameters: number of output layers, sizes of the layers VLD Inverse Quantization Q -1 stream Quantization Q VLC - EL stream BL stream Quantization Q’ VLC

12 Transcoder: New enhancement layers From I to B frames: empty macroblocks can be skipped variable length coding tables of B frames are better suited to encode residual values Advantages: lower importance of base layer size less syntax overhead Both approaches are compliant with new MPEG System proposal

13 Contents Introduction (www) New solutions –System view –Transcoder –Controller –Network sender-receiver –Summary Subjective evaluation Conclusions and future plans

14 Controller: General view Controller chooses how much video data (e.g. layers) should be processed. Takes into account: – amount of available resources (CPU) – amount of video data available (e.g. how many layers have we received) Objective: maximize perceived quality –MAX number of layers to be processed –MIN deadline misses –MIN quality changes Terminal Scalable Video Input Decoder Controller Post-processing Controller

15 Controller: Markov decision process State: progress w.r.t. deadline number of layers decoded maximal number of layers for the next frame Revenue: Reward: number of layers Penalty: deadline misses Penalty: quality change Penalty: quality change, caused be the network

16 Controller: Challenges Processing times -depends on content (stochastic) -depends on layer size Probabilities change for every possible layer size => Unique strategy for every layer size Maximal number of layers for the next frame -network-dependent parameter (stochastic) We need appropriate network behavior model  Unique strategy for every network condition Number of layers in total  Unique strategy for every number of layers

17 Contents Introduction (www) New solutions –System view –Transcoder –Controller –Network sender-receiver –Summary Subjective evaluation Conclusions and future plans

18 Network sender-receiver (NSR) Streams layers over the network Takes care about layer prioritization Sends feedback about receiving conditions Streaming conditions + Receiving conditions = Network conditions μ, BER / PER

19 NSR: Channel Model Using μ, BER / PER we can build a simple channel model [example: Gilbert-Elliott channel (GEC)] We run a simulations for different network conditions and layer configurations For every pair μ, BER / PER we can estimate what is an optimal layer configuration (lookup table) Layer configuration is a run-time input for the transcoder

20 NSR: Controller input Processing times -depends on content (stochastic) -depends on layer size Probabilities change for every possible layer sizelayer size => Unique strategy for every layer size Maximal number of layers for the next frame -network-dependent parameter (stochastic)number of layers We need appropriate network behavior modellayer size  Unique strategy for every network conditionsμ, BER / PER Number of layers in total  Unique strategy for every number of layersnumber of layers BUT! Layer sizes and number of layers have one-to-one relation with μ, BER / PER

21 Contents Introduction (www) New solutions –System view –Transcoder –Controller –Network sender-receiver –Summary Subjective evaluation Conclusions and future plans

22 System: Global view Transcoder Network Sender-Receiver - Terminal - μ, BER / PER number of layers layers sizes strategy Calculated offline Alternative

23 Contents Introduction (www) New solutions Subjective evaluation Conclusions and future plans

24 User tests How does proportion between sizes of BL and EL influence perceived quality? What is better: a large EL with high losses or small EL with low losses? RESULTS: BL size is very important With decrease of overall bit-rate the importance of BL size increases With increase of EL size married to a lower frequency is perceived slightly worse

25 Contents Introduction (www) New solutions Subjective evaluation Conclusions and future plans

26 Conclusions Made a transcoder that produces less bit per EL for the same quality Made a network simulation that allows a better choice of layer configuration Enhanced a terminal controller with realistic network behavior model User tests were performed for perceived quality evaluation of a scalable video

27 Future plans Take into account other terminal resources Enhance the feedback mechanism from the terminal to the source with terminal capabilities information Allow source-terminal resource negotiations Perform subjective test for dynamic behavior of scalable video scheme Implement transcoder on CE platform

28 ?

29 Transcoder: Comparison of the approaches Difference in PSNR between one-layer reference and two-layer scalable coding (the overall bitrate is 5 MBps) I-frame approach B-frame approach

30 NSR: Channel Model Using μ, BER / PER we can build a simple channel model [example: Gilbert-Elliott channel (GEC)] We model the channel as a two-state discrete time Markov chain (DTMC) with states G (good) and B (bad) and four probabilities P, Q, E G, and E B

31 NSR: Channel Model Simulation

32 NSR: Channel Model Outcome For every pair μ, BER / PER we can estimate what is an optimal layer configuration (lookup table) Layer configuration is a run-time input for the transcoder Example: maximal BL bitrate as a function of network conditions