Accomplishing Seeing Time Versatility in Portable Video Spilling Utilizing Adaptable Video Coding - PowerPoint PPT Presentation

achieving viewing time scalability in mobile video streaming using scalable video coding l.
Skip this Video
Loading SlideShow in 5 Seconds..
Accomplishing Seeing Time Versatility in Portable Video Spilling Utilizing Adaptable Video Coding PowerPoint Presentation
Accomplishing Seeing Time Versatility in Portable Video Spilling Utilizing Adaptable Video Coding

play fullscreen
1 / 27
Download Presentation
betsy
Views
Download Presentation

Accomplishing Seeing Time Versatility in Portable Video Spilling Utilizing Adaptable Video Coding

Presentation Transcript

  1. Achieving Viewing Time Scalability in Mobile Video Streaming Using Scalable Video Coding Cheng-Hsin Hsu Senior Research Scientist Deutsche Telekom R&D Lab USA Los Altos, CA Joint Work with Mohamed Hefeeda February 22, 2010

  2. Recent studies show that many users watch videos using their mobile devices even when stationary devices, such as TVs, are available [Vorbau et al., HotMobile’07] Mobile devices have strict power constraint Battery technologies do not grow as fast as CPU speed, memory size, and disk capacity Mobile Video Streaming 2

  3. Mobile users must consider battery lifetme as a new dimension of viewing quality For example: a user watches a 30-min episode Motivation The highest possible quality High quality through-out the episode Until running out of battery Quality Time 0 30 25 Streaming at a lower quality, however, requires transcoding, which is computationally intensive 3

  4. Multiple substreams can be efficiently extracted from a scalable stream for support different resolutions, frame rates, and fidelity levels Often used for devices with heterogeneous resources Such as bandwidth, display resolution, and decoder capability We highlight another benefit of SVC: to enable viewing time scalability Scalable Video Coding (SVC) 4

  5. Goal: allow users to choose desired video quality and viewing time of a streaming video We first develop a framework to predict battery lifetime and perceived quality of individual substreams extracted from a scalable stream We then propose an algorithm and provide users a control knob to trade off perceived quality and viewing time Problem Statement 5

  6. Quality-Power Adaptation Framework Desired Quality (in MOS) Desired Viewing Time (in hr) Framework Perceived Quality Model Comm. Power Model Video Characteristics Optimal Substream Adaptation Algorithm Other Power Model CPU Power Model Device Characteristics Required CPU Cycles Expected Viewing Time (in hr) Expected Quality (in MOS) 6

  7. Divide power consumption into CPU: for video decoding, written as pc(y), where y is number of cycles per second Comm.: for receiving video stream, written as pn(), where is the sleep period ratio Background: accounts for others, such as display backlight, written as a constant pb Power Model 7

  8. Depends on types of wireless networks Such as WLANs, WiMAX, and 3G Cellular We focus on mobile TV networks and propose a broadcast scheme, FMVB, for scalable streams FMVB (Flexible Mobile Video Broadcast) supports substream extractions at mobile receivers Extractions must be done at receivers because of the broadcast nature Enables both temporal (specified by frame rate t) and fidelity (specified by quan. step ) scalability Communication Power Model 8

  9. FMVBdetermines time and size of transmission bursts for individual layers Formulas are in the paper [Theorem] FMVB produces feasible schemes with bursts that are long enough to sustain smooth playout Communication Power Model (cont.) Ch. 1, Layer Ch. 1, Layer Ch. 1, Layer Ch. 2, Layer Ch. 2, Layer Ch. 1, Layer Ch. 2, Layer Ch. 2, Layer Time 9

  10. [Theorem] The sleep period ratio of mobile devices receiving the substream at frame rate and quan. step is: We write , where mW is the power consumption of the network chip Communication Power Model (cont.) Our Comm. Model 10

  11. Complexity model: estimate , which is the number of cycles required by decoding a substream at frame rate and quan. step We can estimate how many cycles needed for each substream, and then? CPU Power Model [Ma and Wang, ICASSP’08]. 11

  12. Modern CPUs can run at a lower frequency to save energy We match produce (CPU frequency) with demand (number of cycles) We adopt the energy scaling model: Combining and functions gives our CPU power model CPU Power Model (cont.) [He et al., CSVT’05]. CPU power efficiency factor 12

  13. Total Power Consumption Battery capacity is often measured in mAh using built-in circuits Viewing time can be written as , where is the battery voltage Total Power Consumption Total CPU Comm. Bg. Number of cycles per second Interface sleep period ratio 13

  14. Quality Model MOS Score 0% -- 100% 75% Perceived Quality Model Full Quality 70% 45% • The perceived quality in MOS is given as: [Wang et al., PV’09]. 14

  15. Leverages on the models to compute expected viewing time (or quality) for each substream Returns the best substream while satisfying users’ requirements Quality-Power Adaptation Algorithms Optimal Substream QPAA Desired Viewing Time Expected Quality Optimal Substream QPAA Desired Quality Expected Viewing Time 15

  16. Energy Saving on Mobile TV Receiver • Use off-the-shelf hardware • Build software in the base station • Transmits DVB-H compliant signals [MM08’Demo] 16

  17. Four model parameters are device dependent : network interface power consumption : network interface transition time : background power consumption : CPU power efficiency factor We demonstrate how to derive the parameters using a Nokia N96 phone Device Dependent Parameter Estimation 17

  18. We encode a 10-min news into a stream at video rate 450 kbps and audio rate 32 kbps We configure our base station to broadcast this stream with different inter-burst periods: 250, 500, 1000, 2000, and 3000 msec We use an phone to watch each broadcast for 3.5 hours, and we measure power consumption 4 times each second We use the measurements to derive mW and msec Network Interface Power Consumption and Transition Time 18

  19. Background Power Consumption and CPU Efficiency Factor • To measure background power consumption, we configure an N96 phone to display pictures in slide show mode for 1.5 hr • Isolate comm. and CPU power consumption • Average bg. power consumption is pb = 290.38 mW • With all other model parameters, we can compute the CPU power consumption, and thus average CPU Efficiency Factor: 19

  20. Tradeoff between Video Quality and Power Consumption • Use an N96 phone with 950 mAh, 3.7 V battery • Configure base station to broadcast at 8.289 Mbps • Consider Crew sequence with frame rates 3.75, 7.5, 15, and 30 fps and quan. steps 16, 40, 64, and 104 • Three broadcast services using the proposed FMVB scheme • Temporal only, quality only, and combined 20

  21. Sample Results • Wide range of viewing time scalability is possible: 4.3 to 11.1 hrs 21

  22. Sample Results (cont.) • Quality degradation is not linear to viewing time increase • Frame rate 30  15 doubles viewing time, but only reduces quality by 5% • Highlights the importance of our framework 22

  23. Sample Results (cont.) • Combined scalability: each point represents a substream • Our framework helps users to pick the best substream 23

  24. Conclusions • Proposed a quality-power adaptation framework to systematically control the tradeoff between video quality and viewing time • Proposed a video broadcast scheme (FMVB) to broadcast scalable streams for viewing time scalability • Presented the estimation of device dependent parameters, and demonstrated the effectiveness of the proposed adaptation framework • Framework can be extended to other wireless networks and applications 24

  25. Questions and Comments 25

  26. Network Interface Power Consumption and Transition Time (cont.) • Longer inter-burst periods  fewer spikes • CDFs indicate that the difference between chip sleep/active modes is mW 26

  27. Network Interface Power Consumption and Transition Time (cont.) • Broadcast schemes with smaller inter-burst periods  more instances • Comparing any two schemes gives us a value • msec, which is aligned with 150 msec reported on a recent DVB-H chip datasheet [Philips] 27