video codec evaluation - universitetet i oslo€¦ · evaluation inf 5080 andreas feuersinger...
TRANSCRIPT
1
Video Codec Evaluation
Inf 5080
Andreas Feuersinger
Overview of presentation
MotivationObjective approaches (algorithms)
PSNR, VQM, SSIMSubjective approaches (human eye)
SAMVIQ, DSIS, DSCQS, SSCQE, SDSCE Expamples
Comparison of evaluationsConclusion
2
Motivation
Acquisition, processing, compression, storage, transmission, reproduction and any variety of distortion which results in a degradation of visual qualityLots of codecs which are all supposed to be the “best”Question: How to measure video quality?
Objective approaches
PSNR (Peak Signal to Noise ratio)“easy”, straight forward approachLogarithmic value of the inverted Mean Square Error
VQM (Video Quality Metric) [2]1991 developed with up to 95% correlation to subjective approachesDividing of temporal an spatial informationA convolution an standard deviation gives then grade of distortionDifferent models (television, videoconferencing, general, ...)
3
Objective approaches - SSIMStructural SIMilarity (SSIM) [1]“The measurement of structural information loss can
provide a good approximation to perceived image distortion.”
Subjective approaches “Golden Device” of video quality is used (human eye)For comparison reasons there are standards defined by ITU (International Telecommunication Union)Standards regulate
strict viewing conditionstype of displays usedsource signal selected15 or more non-expert observerstest session should not last more than half an hourlength of each clip is recommended to be 5 or 10 secondsdefined grading scale
4
Subjective approaches
Standardized by ITU-R BT.500 & ITU-T P.910
DSIS: Double Stimulus Impairment ScaleDSCQS: Double Stimulus Continuous Quality ScaleSSCQE: Single Stimulus Continuous EvaluationSDSCE: Simultaneous Double Stimulus for Continuous EvaluationSAMVIQ: Subjective Assessment Methodology for Video Quality
SAMVIQ [5]Developed by EBU-Group B/VIMMembers:
IRT (Institut für Rundfunktechnik)FTRD (France Telekom R&D)RAI (Radio Audizioni Italiane)NRK (Norsk Rikskringkasting)
Start 2001, Finished 20042001-2002: Development2002-2003: Evaluation of Codecs (CIF,QCIF)2004:Testing
5
SAMVIQ - Motivationprevious methods created for TVdifferences
Features
Featurelist of standards
6
Expamples
CS MSU Graphics&Media Lab Video Group (2006)
Nokia test for handhelds
SC MSU (2006) [6]
DivX 6.0XviD 1.1.0x264 svn-352MWMV 9.0
SAMVIQ as reference for objective measurements
7
CS MSU (2006) testscenes
CS MSU (2006) testscenes
8
CS MSU (2006) results
CS MSU (2006) resultsSubjective against objective graphsSequences must be treated separately Sequence is on a line > good correlation subj. and obj. (like SSIM)
9
Nokia single stimulus method [4]Nokia’s SS test ofH.264/AVC and WMV 9(Mustonen et al, 2003)
H.264 better – especiallyin the lower bitrates
Conclusion
“Golden device” very difficult to model Objective models useful for quick, cost effective testsReliable and accurate results can only be made by “real eyeballs”
10
References[1] http://www.cns.nyu.edu/~lcv/ssim/[2] http://www.its.bldrdoc.gov/n3/video/pdf/vqm_techniques_v2.pdf[3] http://www.ebu.ch/trev_301-samviq.pdf[4] http://people.arcada.fi/~johnny/videocodec.pdf[5] http://www.irt.de/IRT/veranstaltungen/irt-tech-wiss-koll-video-
qualitaet.pdf[6] http://www.compression.ru/video/codec_comparison/pdf/-
msu_subjective_codecs_comparison_en.pdf