Video Capacity of WLANs with a Multiuser Perceptual Quality Constraint Authors: Jing Hu, Sayantan Choudhury, Jerry D. Gibson Presented by: Vishwas Sathyaprakash,

1

Video Capacity of WLANs with a Multiuser Perceptual Quality Constraint

Authors:Jing Hu, Sayantan Choudhury, Jerry D. Gibson

Presented by:Vishwas Sathyaprakash, Aditya Sharma

2

Overview

• Introduction & Motivation• Experiment – Simulation Setup• Results of Simulation– Packet Loss and Video Quality– Video Data Rate

• Suggested Improvements• Determining Video Capacity• Applications• Conclusion

3

Introduction

• Wireless LANs gaining popularity• Multimedia - Large part of WLAN traffic• How many users can be supported ?• Video Quality - as perceived by users• Fine Balance - Capacity v/s Quality• Authors Define this fine balance in terms of

perceived quality of the video being delivered to r% of the users– Example: Streaming a video in this class

4

Motivation

• Video Compression Variable Size & Quality• Measuring video quality: – Mean Squared Error & Peak Signal-to-Noise Ratio– Poor co-relation to perceived video quality– HVS : Computationally Expensive

• Capacity, Encoding and transfer rates– Capacity calculation not defined clearly– Transfer rate/Capacity depends on codec used

• First effort to relate Quality and Capacity: Not been studied yet

5

Overview

Introduction & MotivationExperiment – Simulation Setup• Results of Simulation– Packet Loss and Video Quality– Video Data Rate


6

Video Over WLAN: Simulation Setup

• Video Codec: H.264– Packetized Video; Many coding schemes/options

• GOPs of 10, 15, 30, 45• 3 Videos: 90 Frames each– Silent.cif; Paris.cif; Stefan.cif

7

Video Over WLAN: Simulation Setup

• WLAN: IEEE 802.11a (5GHz; 54 Mbps)• Quantization Parameters (QP): 26 (fine) & 30

(coarse)• Payload Size (PS) of 100 and 1100 bytes• Noise: Additive White Gaussian Noise (AWGN)• Packet Loss Compensation: Base Model– I-Frame: Recovery of MB by Spatial Interpolation– P-Frame: Copying MB from reference frames– Lost Frame: Entire frame is copied

• Measurements: SNR, PER and Data Rates (DR)

8

Overview

Introduction & MotivationExperiment – Simulation SetupResults of Simulation

Packet Loss and Video Quality Video Data Rate


9

Results of Simulation: Packet Loss (1)N

umbe

r of R

ealiz

ation

s

10

• CDFs of PER > 0 at 0 realizations• CDFs of PER < 1 for realizations <1 • Average PER over realizations of multipath fading not

an appropriate indicator of channe performance• Variation of PER of AWGN channel is less, ranges from

1% to 3%• Avg. PER of multipath channels = 5.5% This

represents only a small number of total realizations • Avg. PER of AWGN channel is much lower than Avg.

PER of channels with fading

Results of Simulation: Packet Loss (2)

11

Results of Simulation: Packet Loss (3)

70%: No Packet Loss

90%: Packet Loss < 2%

Num

ber o

f Rea

lizati

ons

12

Results of Simulation: Packet Loss (3)…contd.

• Video: Silent.cif• QP = 26, 30• GOP = 15• PS = 100• Fading Channels• Thick lines represent

Average PSNR• ‘+’ marks: 70% of the

overlapping realizations with no packet loss

Frame Index; QP = 26

PSN

R of

eac

h fr

ame

PSN

R of

eac

h fr

ame


13

• Video: Silent.cif• QP = 26, 30• GOP = 15• PS = 100• Shows behavior of

noise channels (AWGN)

• Prediction in video encoding causes realizations with similar PER: yet, completely different video quality

Results of Simulation: Packet Loss (4)PS

NR

of e

ach

fram

ePS

NR

of e

ach

fram

e



14

Results of Simulation: Data Rate (1)

15

Results of Simulation: Data Rate (2)

16

Overview


Packet Loss and Video QualityVideo Data Rate

Suggested Improvements• Determining Video Capacity• Applications• Conclusion

17

PSNRr,f & Perceptual Quality of Multiple Users MOSr

• Defined as the PSNR achieved by f% of the frames in each one of the r% realizations

• f%: Captures the majority of the frames• r%: Captures the reliability of a channel over many users • Claim: Quality Perception doesn’t change for high ‘f’• Observations behind the claim:

– Poor quality frames dominate viewers’ experience– Quality drop in a very small number of frames is not perceivable

by the human viewer– PSNR > threshold Increase in PSNR doesn’t translate to

increase in perceptual quality

18

PSNRf and Perceived Video Quality (1)

• Experiment to prove the claim:– Same video sequences played side-by-side– Left: Raw video / Perfect Quality– Right: Compressed Video with recovered packet

losses and concealment– 3 humans: Rate the videos from 0 – 100%– Scores plotted

19


20

• Of all f values, f=90% correlates to best opinion score• Mean Opinion Score achieved by r% of the transmissions is

given by:

• Dotted Lines = Average PSNR (existing)• Problem with regular PSNR:

– Quantitative measure of Quality– Underestimates the quality at high quality levels– Overestimates the quality at low quality levels

• Thick lines = PSNRf (proposed)– Serves as effective quality measure: correctly estimates low quality

and high quality as perceived by HVS


21

Overview



Suggested ImprovementsDetermining Video Capacity• Applications• Conclusion

22

Video Capacity of WLAN with DCF (1)

• Thumb-rule: Video frames must arrive at the buffers before the playback deadline.

• DCF: Distributed Coordination Function: based on CSMA/CA

• Requirement to know the number of users supported:– Network operators get an idea of number of users that

can be supported for identical traffic (capacity planning)– Mix of users having different traffic demands capacity

is approximated to an interpolation of capacity values for each traffic category

23

• Video capacity with no buffer at receiver– I-Frame and P-Frame sizes differ greatly– What happens when:• All users are transmitting I-Frame? - Worst Case

• All users coordinate I-Frame transmission ? - Best Case


24

• Video capacity with play-out buffer at receiver (Cb)– Play-out buffer b is the length of the buffer (ms)– It is used only for the frames that have more bits

than the other frames (I-Frame)– How buffer length compares with SI/SP : Plot

– Cb fluctuates between CM and a lower bound given by:


25

• Length of buffer required for video capacity to reach upper bound, for typical SI/SP values:


26

Overview



Suggested ImprovementsDetermining Video CapacityApplications• Conclusion

27

Quality Constrained Video Capacity and its Applications (1)

28


29

• Observations:– Users watching silent.cif Excellent Quality– Users watching paris.cif Average Quality– Users watching stefan.cif Poor Quality

• Applications:– Link Adaptation based on capacity– System performance evaluation– Accurate System Design


30

Overview



Suggested ImprovementsDetermining Video CapacityApplicationsConclusion

31

Conclusion

• Average PER / Average PSNR Not a suitable indicator of video quality

• They should not serve as the basis for video quality assessment

• Proposed ‘perceptual’ quality indicator matches the quality with the human vision system’s quality perception

• Video Capacity with/without buffering• Quality Indicator + Video Capacity design better WLAN

communication system with importance to both quality and efficient capacity utilization

32

Conclusion (2)

• Some observations:– Video Quality Perception is a highly subjective test – Only 3 human ‘testers’ considered: Results could vary with more

humans testing the theory– Ideal test conditions assumed:

• Packets/Frames received without errors• Collisions are not considered• Single Hop networks considered. More Packet errors / losses in multi-hop

networks: Losses affect video quality directly

– 802.11a has been used for testing; Widespread use of 802.11g/n today: Multipath Fading parameters may change due to operating frequency in 802.11g/n (2.4GHz) + Interference

– Little description of actual applications of the proposed method– Portability of proposed method assumed: Same results expected in

other video coding methods (MPEG-2) but not proved

Q&A33Note: All images are the property of the respective owners. Images used for non-profit / educational purposes only.

Video Capacity of WLANs with a Multiuser Perceptual Quality Constraint Authors: Jing Hu, Sayantan Choudhury, Jerry D. Gibson Presented by: Vishwas Sathyaprakash,

Documents