tocModeling Packet-Loss Visibility in MPEG-2 VideoSandeep Kanumuri, Student Member, IEEE, Pamela C. Cosman, SeniorI. I NTRODUCTIONFig.€1. Illustration of FR, RR, and NR methods.II. R ELATED W ORKIII. E FFECT OF A P ACKET L OSSIV. S UBJECTIVE T ESTSFig.€2. Histogram of response times.V. S TATISTICAL B ACKGROUND CART AND GLMFig.€3. Histogram of times between adjacent losses.A. CARTB. GLMVI. F ACTORS A FFECTING V ISIBILITYFig.€4. FRAMETYPE value for different frames in a GOP.TABLE I D ESCRIPTION OF F ACTORS A FFECTING V ISIBILITYVII. CART R ESULTSVIII. L OGISTIC R EGRESSION R ESULTSFig.€5. Comparison of classification accuracy. For example, the Fig.€6. CART classifier tree in the NR-B case.TABLE II C OEFFICIENTS FOR M ODEL 3 IN NR-BA. Other ModelsIX. GLM FOR C LASSIFICATIONA. Classification Based on Probability of VisibilityFig.€7. Factor Significance: Plot showing increase in deviance tB. Comparison CART and GLMFig.€8. Plot of Deviance for models considered.Fig.€9. NR-B: Cross-validation accuracy versus $\alpha$ for GLM Fig.€10. GLM classifiers: Comparision of RR, NR-P, and NR-B methFig.€11. GLM classifiers: Number of decisions made versus $\alphX. C ONCLUSIONFig.€12. Comparison: Classification accuracy of GLM and CART basA. R. Reibman, V. Vaishampayan, and Y. Sermadevi, Quality monitoM. Masry and S. Hemami, A metric for continuous quality evaluatiP. Gastaldo, S. Rovetta, and R. Zunino, Objective quality assessO. Verscheure, P. Frossard, and M. Hamdi, Joint impact of MPEG-2J. Lu, M. Chatterjee, M. D. Schwartz, M. K. Ravel, and W. M. OsbA. E. Conway and Y. Zhu, Applying objective perceptual quality aG. W. Cermak, Videoconferencing Service Quality as a Function ofB. Chen and J. Francis, Multimedia performance evaluation, AT&T S. Mohamed and G. Rubino, A study of real-time packet video qualC. J. Hughes, M. Ghanbari, D. E. Pearson, V. Seferidis, and J. XS. Winkler and R. Campos, Video quality evaluation for internet K. Brunnstrom and B. N. Schenkman, Quality of video affected by A. Watson and M. A. Sasse, Measuring perceived quality of speechM. S. Moore, J. M. Foley, and S. K. Mitra, Detectability and annM. S. Moore, S. K. Mitra, and J. M. Foley, Defect visibility andM. G. Ramos and S. S. Hemami, Suprathreshold wavelet coefficientD. Chandler and S. S. Hemami, Effects of natural images on the dZ. Yu, H. R. Wu, S. Winkler, and T. Chen, Vision-model-based impS. Wolf and M. Pinson, In-Service Performance Metrics for MPEG-2L. Breiman, J. Friedman, R. Olshen, and C. Stone, ClassificationP. McCullagh and J. A. Nelder, Generalized Linear Models, 2nd edA. R. Reibman, S. Kanumuri, V. Vaishampayan, and P. C. Cosman, VS. Kanumuri, P. C. Cosman, and A. R. Reibman, A generalized lineIEEE TRANSACTIONS ON MULTIMEDIA, VOL. 8, NO. 2, APRIL 2006 341Modeling Packet-Loss Visibility in MPEG-2 VideoSandeep Kanumuri, Student Member, IEEE, Pamela C. Cosman, Senior Member, IEEE,Amy R. Reibman, Senior Member, IEEE, and Vinay A. Vaishampayan, Senior Member, IEEEAbstract—We consider the problem of predicting packet loss vis-ibility in MPEG-2 video. We use two modeling approaches: CARTand GLM. The former classifies each packet loss as visible or not;the latter predicts the probability that a packet loss is visible. Foreach modeling approach, we develop three methods, which differ inthe amount of information available to them. A reduced referencemethod has access to limited information based on the video at theencoder’s side and has access to the video at the decoder’s side. Ano-reference pixel-based method has access to the video at the de-coder’s side but lacks access to information at the encoder’s side.A no-reference bitstream-based method does not have access to thedecoded video either; it has access only to the compressed video bit-stream, potentially affected by packet losses. We design our modelsusing the results of a subjective test based on 1080 packet losses in72 minutes of video.Index Terms—Packet-loss visibility, perceptual quality metrics,subjective testing, video quality.I. INTRODUCTIONWHEN sending compressed video across today’s commu-nication networks, packet losses may occur. Networkservice providers would like to provision their network to keepthe packet loss rate below an acceptable level, and monitor thetraffic on their network to assure continued acceptable videoquality. Traditional approaches to video quality assume that allpacket losses affect quality equally. In reality, packet losses havedifferent visual impacts. For example, one may last for a singleframe while another may last for many; one may occur in themidst of an active scene while another is in a motionless area.Not all such packet losses are visible to the average humanviewer. Thus, the problem of evaluating video quality givenpacket losses is challenging. As a first step toward developinga quality metric for video affected by packet losses, we addressthe problem of packet loss visibility in our current work.Our long-term goal is to develop a quality monitor that isaccurate, real-time, can operate on every stream in the net-work and answers the question, “How are the losses presentin this particular stream impacting its visual quality?” In thisManuscript received October 21, 2004; revised May 20, 2005. This work wassupported in part by the National Science Foundation, the Center for WirelessCommunications at UCSD and the UC Discovery Program of the State of Cal-ifornia. The associate editor coordinating the review of this manuscript and ap-proving it for publication was Dr. Pascal Frossard.S. Kanumuri and P. C. Cosman are with the Department of Elec-trical and Computer Engineering, University of California-San Diego(UCSD), La Jolla, CA 92093-0407 USA (e-mail: [email protected];[email protected]).A. R. Reibman and V. Vaishampayan are with AT&T Labs—Research,Florham Park, NJ 07932-0971 USA (e-mail: [email protected]; [email protected]).Digital Object Identifier 10.1109/TMM.2005.864343paper, we focus on predicting the visibility of packet lossesin MPEG-2 compressed video streams. Toward this goal, wedevelop statistical models to predict the visibility of a packetloss. We use a well known statistical tool called Classificationand Regression Trees (CART) to classify each packet loss asvisible or invisible. We use a generalized linear model (GLM)to predict the probability that a packet loss will be visible toan average viewer. The input to these models consists of
or
We will never post anything without your permission.
Don't have an account? Sign up