References[1] Ramanathan Palaniappan, Nitin Suresh and Nikil Jayant, “Objective measurement of transcoded video quality in mobile applications”,IEEE MoVID 2008 (workshop as a part of WoWMoM 2008), Newport Beach, CA, June 2008.
Conclusion & Current Work
• MTBF estimates from AVQ scores :– show a wide range across transcoding bit rates and codecs.– are subjectively more meaningful.– better represent the slight variations in degraded visual quality.
• Our current focus (in a Cisco Research Project) : – the analysis of network artifacts (NA) on video quality– Using these automatic VQ measurements to enhance the streaming of IP video
Today’s Video Delivery Scenario
The Video Quality (VQ) Challenge
• Develop subjectively meaningful metrics
• Must be objective, enabling real time computation
• Zero Reference (ZR) nature – independent of source video
Our Solution
• Mean Time Between failures (MTBF) – Failure refers to video artifacts deemed to be perceptually noticeable– Directly related to Mean Opinion Score (MOS)
• Automatic Video Quality (AVQ)– Objective estimate of MTBF for the transcoding experiments
Metrics used in our work
• PSNR – Simple full Reference metric– Computes pixel by pixel difference per frame basis b/w source and processed video– Actual value not definitive but comparison b/w two values measures quality
• Automatic Video Quality (AVQ)– Zero Reference metric– Based on Spatio temporal algorithms and knowledge of Human Visual System
AVQ Metric
• Computation based on output pixel values• AVQ score shows excellent subjective attributes
Our Transcoding Platform – VLC
• Flexible transcoding options like different codecs, variable bit rate, GOP size etc.
• 2 transcoding operations– MPEG 2 to H.264 – MPEG 2 to MPEG 4 SP
Experimental Setup
Results
Sample Frames
Objective Measurement of Transcoded Video Quality in Mobile Applications
Ramanathan Palaniappan, Nitin Suresh and Nikil JayantSchool of ECE, Georgia Institute of Technology, Atlanta, GA 30332
Network
Video Server & Transcoder
Projectors
3G Smart Phones
HD TVs
LaptopsLCD Displays
MPEG 2 to H.264Transcoder
384 – 192 Kbps
MPEG 2 to MPEG 4Transcoder
384 – 192 Kbps
MPEG 2 Transrater384 – 192 Kbps
MPEG 2 512 Kbps72 sec (QCIF)
H.264
MPEG 4
MPEG 2
0 500 1000 150020
30
40
50
60
70PSNR vs frame number @ 192 Kbps
Frame number
PS
NR
(dB
)
H.264MPEG 4MPEG 2
0 500 1000 15000
0.1
0.2
0.3
0.4
0.5
0.6
0.7AVQ CA score vs frame number @ 192 Kbps
Frame number
AV
Q C
A s
core
H.264MPEG 4MPEG 2
0 500 1000 15000
0.1
0.2
0.3
0.4
0.5
0.6
0.7AVQ CA score vs frame number for MPEG 4
Frame numberA
VQ
CA
sco
re
384 Kbps256 Kbps192 Kbps
0 500 1000 150025
30
35
40
45
50PSNR vs frame number for MPEG 4
Frame number
PS
NR
384 Kbps256 Kbps192 Kbps
Fig. 1 : (a) plots PSNR for the video transcoded into H.264, MPEG 4 and MPEG 2 at 192 Kbps. (b) plots AVQ compression artifact (CA) score for the same case (Range : 0 – best and 1 – worst).
(a) (b)
Fig. 2 : (a) plots PSNR for the same video transcoded at 192, 256 and 384 Kbps with the MPEG 4 transcoder. (b) plots AVQ compression artifact (CA) score for the same case.
(a) (b)
Bit rate 192 kbps 256 Kbps 384 Kbps
Codec MTBF PSNR MTBF PSNR MTBF PSNR
H.264 546 36.51 1406 37.31 10000 38.24
MPEG-4 7 34.37 22 35.65 1968 37.39
MPEG-2 5 33.90 13 34.88 140 36.78
Table 1 : PSNR (dB) & estimated MTBF (sec) based on AVQ CA Score
(a) (c) (b)
Fig. 3 : Frame # 508 transcoded at 192 Kbps into (a) MPEG 2, (b) MPEG 4 & (c) H.264
Fig. 4 : Frame # 1469 transcoded into MPEG 4 at (a) 192 Kbps, (b) 256 Kbps & (c) 384 Kbps
(a) (b) (c)