F Smooth Extracton of SVC Fne-Granular SNR Scalable Vdeos wth a Vrtual-GOP-Based Rate-Dstorton Modelng * Jun a SunF, Wen Gao a, Debn Zhao b a Insttute of Dgtal Meda, EECS, Pekng Unversty, Bejng, 1871, Chna; b Department of Computer Scence, Harbn Insttute of Technology, Harbn, 151, Chna ABSTRACT Fne-Granular SNR scalable FGS) technologes n H.6/AVC-based scalable vdeo codng SVC) provde a flexble and effectve foundaton for scalng FGS enhancement layer EL) to accommodate dfferent and varable network capactes. To support smooth qualty extracton of SVC FGS vdeos, t s mportant to obtan the Rate-Dstorton R-D) functon of each pcture or group of pctures GOP). In ths paper, frstly, we ntroduce the R-D analyss of SVC FGS codng n our pror work. Then, wth the analyss and models, we present vrtual GOP concept and a vrtual-gop-based packet schedulng algorthm s proposed to acqure the optmal packet schedulng sequence n a vrtual GOP. Based on the packet schedulng algorthm and the R-D analyss of FGS EL, an effectve and flexble D-R model s proposed to descrbe the D-R functon of the vrtual GOP. Then, wth the R-D model of vrtual GOPs, a practcal non-search algorthm for smooth qualty reconstructon s ntroduced. Compared to the qualty layer method, the reconstructed vdeo qualty s mproved not only objectvely but also subjectvely. Keywords: Smooth qualty, rate-dstorton analyss, scalable vdeo codng SVC), vdeo codng 1. INTRODUCTION Internet s experencng explosve growth of vdeo streamng. Snce the Internet s a shared envronment, t s desrable to encode vdeo wth fne-granular SNR scalable FGS) technologes, so that t can be encoded once, but transmtted and reconstructed many tmes at dfferent targetng rates. The developng scalable vdeo codng SVC) standardzaton project [1] chooses scalable extenson of H.6/AVC as a start pont, whch realzes the fne-granular SNR scalablty through sub-btplane-based progressve refnement of the FGS enhancement layer EL). Snce the gap between SVC FGS scheme and sngle-layer codng can be qute small, the codng scheme ganed sgnfcant nterest. Note that FGS codng mode has been taken away from SVC fnal amendment, a phase SVC project s started, whch may nclude FGS codng mode []. To best utlze SVC FGS vdeos, a bt-stream extracton rate allocaton) algorthm should be employed to optmally transfer the targetng bt rate nto the rate assgned to each FGS pcture. Typcally, there are two optmzaton goals. The frst s optmal extracton wth rate dstorton sense, whch mnmzes the average dstorton subject to the rate constrant [3]. The second s smooth qualty extracton, whch expects to acheve constant qualty wthn the constrant of targetng bt rate. Psychologcal research ndcates that the human vsual system prefers a vdeo sequence havng consstent vsual qualty []. In fact, a vdeo sequence wth hgher average PSNR, but larger varaton n vsual qualty, s generally consdered nferor to a vdeo sequence wth lower average PSNR, but smaller varaton. In ths paper, snce the cascadng quantzaton s a nce method n a GOP [1], we propose smooth reconstructon between dfferent GOPs and optmal extracton n a GOP. Frstly, we ntroduce the R-D analyss and models of scalable vdeo codng n the authors pror work. Then, wth above R-D models, vrtual GOP concept s ntroduced and a vrtual-gopbased packet schedulng algorthm s proposed to acqure the optmal packet schedulng sequence n a vrtual GOP. Based on the packet schedulng algorthm and the R-D analyss of FGS EL, an effectve and flexble D-R model s proposed to descrbe the D-R functon of vrtual GOP. Fnally, usng the vrtual-gop-based R-D model, a practcal nonsearch algorthm s proposed for smooth qualty reconstructon. Compared to the qualty layer method, the reconstructed vdeo qualty s mproved not only objectvely but also subjectvely. * Ths work was supported by Natonal Key Technology R&D Program under contract No. 6BAHA13 and Chna Postdoctoral Scence Foundaton 63.
. RATE-DISTORTION ANALYSIS OF SVC FGS CODING.1 Rate-Dstorton Analyss and Modelng of SVC FGS EL In our pror work of [5], we analyze the D-R functon of SVC FGS EL wth generalzed Gaussan model and conclude the D-R functon wth PSNR crteron should be the left half of a concave functon as a whole. After that, n [6], we further examne the sub-btplane technology of SVC FGS codng and conclude that the D-R functon wth mean square error MSE) crteron should be lnear wthn a FGS layer. Consderng the above two propertes of SVC FGS codng, a pecewse lnear model s proposed to descrbe the R-D functon of FGS EL. Fg. 1 shows the dstorton-rate functons and ts models n the Foreman sequence. PSNR [db] Rate-dstorton comparson of actual FGS codng and GGD-based analyss PSNR crteron) 8 7 6 5 Frame codng 3 Frame analyss Frame codng 1 Frame analyss 39 38 37...6.8 1 1. 1. 1.6 bts/sample MSE 1 1 1 8 6 Rate-dstorton comparson of actual FGS codng and lnear analyss MSE crteron) Frame codng Frame analyss Frame codng Frame analyss...6.8 1 1. 1. 1.6 bts/sample a) PSNR-based dstorton-rate functon b) MSE-based dstorton-rate functon Fg. 1. The dstorton-rate functons and ts models wth PSNR crteron a) and MSE crteron b) n the Foreman sequence. Encoder precdton pont Orgnal value Orgnal value Resdue value ε A ε Ael ε H PreDrft ε C ε Cel Enhancement layer Decoder predcton Decoder predcton Base layer A H C Fg.. The predcton structures of SVC herarchcal B codng.. Dstorton Modelng of Reconstructed Frame wth Drft In the closed-loop codng wth herarchcal B pctures, the current frame dstorton s dependent on the encoder-decoder msmatch of reference frames and the reconstructed resdue frame. Fg. shows the typcal predcton relaton between the current frame and the reference frames usng a pxel-based analyss. ε A and ε C show the dfferences between the orgnal values and the reconstructed values n the reference frames A and C, respectvely. ε A el and ε C el show the correspondng encoder-decoder msmatch values of the reference frames A and C, respectvely. ε H shows the dfference between the orgnal resdue value and the reconstructed resdue value n the current resdue frame H. The pre-drft n Fg. represents the drft of frame A caused by reference frames of A. All varables ε A, ε C, ε A el, ε C el, and ε H may be postve or negatve.
In [6], usng a pxel-based analyss, a smple frame-based dstorton model s proposed to estmate the reconstructed frame dstorton wth drft: 1 1 E ε ) = E ε ) + p ) E ε ) + p ) E ε ) + μe ε ) E ε ) + ν 1) where C ) B) B H fwd Bwd A bwd Bwd C A C E ε s the reconstructon dstorton of target frame B, E ε s the dstorton of FGS EL n frame B, H ) E ε and E ε are the reconstructon dstorton of reference frames A and C, and p fwd, p bwd, and p bwd denote the fracton of forward-predcted, backward-predcted, and b-predcted pxels n frame B. μ and ν are two parameter, whch could be estmated usng two decodng passes such as base layer extracton and full stream extracton. Wth the frame-based drft dstorton model and the pecewse lnear R-D model of FGS EL, we can estmate the reconstructon dstorton of each frame just based on the allocated bt rate of each pcture. A) Vrtual GOP VGOP) Group of pctures GOP) Halved sze f n two VGOPs Vrtual GOP VGOP) Group of pctures GOP) Halved sze f n two VGOPs 1 3 5 6 7 8 1 3 5 6 7 8 Fg. 3. Vrtual GOP defnton and the optmal schedulng algorthm. 3. VIRTUAL-GOP-BASED PACKET SCHEDULING AND RATE-DISTORTION MODELING 3.1 Vrtual-GOP-based optmal packet schedulng algorthm From a gven FGS-coded btstream and a specfc average bt rate, there are many dfferent ways to extract a subbtstream. The same target bt rate could be acheved by selectng and truncatng dfferent FGS levels or packets). However, the codng effcency s sgnfcantly nfluenced by the extracton scheme. Typcally, the FGS levels should be assgned dfferent prorty accordng to the prortzaton order, whch s determned by the D-R slope of each FGS level. However t s not trval to acqure the specfc D-R slope of an FGS level snce t s not only dependent on the propertes of the FGS level, but also dependent on the defnte schedulng sequence of the related FGS levels. For a GOP sze 8 and FGS layers, there need at least 8+7+6+5++3++1)*=7 testng arrangements to acheve the optmal order even just n a GOP, not consderng the nterference of dfferent GOPs. So t s not easy to acqure the optmal schedulng order n a vdeo sequence. Accordng to the herarchcal B structure, the drft of FGS levels propagates wthn a GOP except the levels of the key pctures that could use the base layer as reference. So for smplcty of the prorty settng n the entre btstream, the optmal prorty settng algorthm of ths paper s confned n a GOP and the key pcture levels s consdered n a specal way. The new concept of vrtual GOP s proposed to show the target scheduled pctures and the specal treatments of key pctures. Fg. 3 shows the vrtual GOP defnton n ths paper, where the vrtual GOP contans two consecutve key pctures and all the B pctures between the two key pctures. For the two-layers FGS codng wth GOP sze 8 n Fg. 3, the vrtual GOP sze s 9, and the number of FGS levels that should be consdered for prorty settng s *9=18. The FGS level sze of a key pcture that belongs to two vrtual GOPs s halved n the vrtual GOP. The optmal prorty settng algorthm n a vrtual GOP s descrbed as follow: 1) Calculate the pecewse lnear model of each frame durng the FGS encodng process as n Secton.1. That s, H s derved, whch represents the quantzaton error of resdue frame at FGS level l, where l= l
can represents the quantzaton error of resdue frame n the base layer. The parameters p fwd,, p bwd, and p Bwd, of frame n Equaton 1) are also collected durng the encodng process. ) Estmate the two parameters μ and ν of Equaton 1) n each frame. Frst, two extracton patterns of the base layer extracton and full layers extracton are selected and decoded to acqure the MSE of each frame. Then wth the data example, let b E and p, p bwd, and p Bwd, n step 1), the two parameters μ and ν are calculated. For l H, fwd, f E represent the decoded MSE of frame n the base layer extracton and full layers ν of frame n HFg. 3H be calculated by extracton respectvely. Then the drft model parameters μ and the followng two equatons: b 1 b 1 b b b E = H + pfwd, Bwd,) E + pbwd, Bwd,) E8 + μee8 + ν. f 1 f 1 f f f E = H + pfwd, Bwd,) E + pbwd, Bwd,) E8 + μe E8 + ν 3) Set the prorty order and obtan the correspondng PSNR-rate slopes of FGS levels n each vrtual GOP: For each vrtual GOP n the btstream A) Intalze the MSE array of the vrtual GOP MSE VGOP [1..gop_sze+1)] as the reconstructon b b b b b b b b b MSE of the base layer, whch s [ E, E1, E, E3, E, E5, E6, E7, E 8] for the frst VGOP Results could be derved drectly n step )). Intalze the level array as the 1st FGS level n the vrtual GOP. In HFg. 3H, the level array could be descrbed as array_level =[,1), 1,1),,1), 3,1),,1), 5,1), 6,1), 7,1), 8,1)] the green crcle packets n HFg. 3H), where *,*) represents frame_no, FGS_level_no). B) Select contnuously the best FGS level wth the largest PSNR-rate slope to schedule untl the array_level s NULL: ) Calculate the dstorton decrease of each FGS level n current array_level. In HFg. 3H, for example, f the FGS level,1) s tested and added, the MSE decrease of frame s frst 1 derved that s decmse,1) [] = H H, decmse,1) [ k ] = ; ). After that, from temporal level to temporal level 3, the MSE decrease of frame, 6, 1, 3, 5, 7 wth FGS level,1) are calculated orderly usng HEquaton 1)H. For example, the MSE decrease of frame wth FGS level,1) could be calculated as 1 decmse,1) [] = MSEVGOP[] H + p fwd, Bwd,) MSEVGOP [] decmse,1) []) 1 + pbwd, Bwd,) MSEVGOP [] decmse,1) []) + μ MSEVGOP[] decmse,1) []) MSEVGOP[] decmse,1) []) + ν ) After the process, the dstorton decrease decmse,1) [.. gop_sze+1)] of the FGS level,1) n current vrtual GOP could be derved, and so the PSNR ncrease of the FGS level,1) could be calculated as : ncpsnr,1) = 1 meanlog 1 MSEVGOP[1.. gop _ sze + 1)])) 1 meanlog 1 MSEVGOP[1.. gop _ sze + 1)] decmse,1) [1.. gop _ sze + 1)])) ) Calculate the PSNR-rate slope of each FGS level n the current array_level, and then select the FGS level wth the largest PSNR-rate slope to schedule and output ts correspondng PSNR-rate nformaton. ) Update array_level and MSE VGOP : The optmal FGS level n ths cycle s removed from array_level and the hgher level value of the correspondng frame s loaded n array_level. MSE VGOP s updated wth the optmal FGS level added. For example, n HFg. 3H, f,1) s the best level of ths cycle, now array_level s set as [,1), 1,1),,1), 3,1),,), 5,1), 6,1), 7,1), 8,1)] the green crcle packets n the rght of HFg. 3H), and MSE VGOP =MSE VGOP - decmse,1). ) Calculate the D-R slope of each FGS level: The level D-R slope of the key pctures n two vrtual GOP s averaged, and the D-R slope of other packets s same as the one n the vrtual GOPs. In the above algorthm, two sgnfcant data are obtaned. The frst s the prorty order of FGS levels n each vrtual GOP, the other s the correspondng D-R nformaton of the vrtual GOP. The prorty order and D-R nformaton of the vrtual GOPs s modeled and used for smooth qualty reconstructon.
3. Rate-Dstorton Modelng of Vrtual GOP In Secton.1, we conclude that the dervatve of D-R functon PSNR crteron) n FGS EL decreases as bt rate ncreases, that s, the D-R slope of lower-layer FGS packet should be steeper than the hgher-layer FGS packet n the same pcture. And n the algorthm of Secton 3.1, the lower-layer packet s loaded frst to the array_packet, and the packet wth the hghest D-R slope n the array_packet s always scheduled frst untl all the packets n the vrtual GOP are scheduled. So, generally, we can assume that the dervatve of the D-R functon PSNR crteron) n a vrtual GOP decreases contnuously as the rate ncreases. Then, as a smlar method of our prevous work n [7], the complete formula of the proposed R-D model n a vrtual GOP can be descrbed as follows: PSNR R) = a * R + A A B)/1 + b* R). ) where R s the bt rate bt/sample), B s average PSNR of base layer n a vrtual GOP, a, A and b are the control parameter. Usually, a and b could also be selected as constants for a coarse approxmaton. To verfy the vrtual-gop-based R-D model, three experments are desgned. In the frst experment, a s set as the constant 6. and b s set as the constant 8.. The parameter A s acqured by the nonlnear least-squares data fttng. In the second experment, b s set as the constant 8., and the other two parameters a and A) of the asymptote are calculated usng the same nonlnear least-squares data fttng. In the thrd experment, all the three parameters a, b and A) of the R- D model are acqured through the nonlnear least-squares data fttng. Table 1 shows the means of the maxmum and average absolute estmate errors across all vrtual GOPs n the eght test sequences. It s observed that the average absolute estmate error of the complete model on all the test vdeos s only about.37 db n the thrd experment. Generally, the proposed R-D model s accurate and flexble, and usually, two-control-parameter model s good enough to balance the accuracy and the complexty of the R-D model. Table 1. The Average of Maxmum and Average Absolute Estmate Error across the Vrtual GOPs Average across all GOPs Foreman Moble Bus Football Exp1 Exp Exp3 Exp1 Exp Exp3 Exp1 Exp Exp3 Exp1 Exp Exp3 Max error db).18.151.17.19.16.16.68.19.9.7.1.63 Ave error db).75.58.6..5.5.9.38.9.1.6.3 Average across Cty Crew Harbour Soccer all GOPs Exp1 Exp Exp3 Exp1 Exp Exp3 Exp1 Exp Exp3 Exp1 Exp Exp3 Max error db).1.1.179.397.117.17.616.7.113.33.13.9 Ave error db).5.9.7.19.1.38.385.57.3.63..3. SIMPLE ALGORITHM FOR SMOOTH QUALITY RECONSTRUCTION SASQR).1 Smple Algorthm for Smooth Qualty Reconstructon SASQR) In SVC codng, although the cascadng of the quantzaton parameters over herarchy levels results n relatvely large PSNR fluctuatons nsde a group of pctures, subjectvely, the reconstructed vdeo appears to be smooth [1]. So t s nce to smooth the vdeo qualty between dfferent GOPs. In the envronments of vrtual GOP model, the key problem s how to truncate vrtual GOPs to both match the avalable average bandwdth R and acheve a constant qualty D targ et for each vrtual GOP. To obtan the D, we need to calculate a combned functon CD ), whch s constraned by R : targ et N 1 1 CD ) = R D) = R N, 3) = where R D) s the R-D functon of vrtual GOP, N s the number of smoothed vrtual GOPs. Through 3), we can 1 obtan the target constant qualty Dtarg et = C R). Then the allocated bt rate of vrtual GOP can be calculated by 1 1 R Dtarg et) = R C R)). However t s dffcult to get a closed-form soluton of C for the known R-D models and a search algorthm for D targ et s a burden for streamng server snce the R changes contnually n the actual streamng. Usng the R-D model of vrtual GOPs, where b s fxed as 8., the algorthm SASQR s descrbed as follows: 1) Calculate the average dstorton D wth unform bt rate allocaton R :
N 1 N 1 N 1 N 1 [ * )/1 8.* )]/ = = = =, ) D = R a + A A B + R N where a, A and B are the correspondng values of ) n vrtual GOP. ) Calculate the ntal bt rate allocaton nt _ rate = R D), where R D ) s the nverse functon of H)H n vrtual GOP. Then we calculate the average tune bt rate: N 1 1 tune _ rate = nt _ rate ) R. 5) N = 3) Calculate the tune weght of each vrtual GOP : 1 ' 1 ' 1 tune _ weght = N *[ D nt _ rate )] [ D nt _ rate )], 6) / N = where ' ' 1 ' D nt _ rate ) s the dervatve of the D-R functon at the btrate nt_rate. [ D nt _ rate)] = R D) approxmates the btrate requrement of one unt dstorton change at the dstorton pont D of vrtual GOP. ) Calculate the transmttng bt rate n a vrtual GOP by trans _ rate = nt _ rate tune _ rate* tune _ weght. 7) After acqurng tran _ rate of each vrtual GOP, the bt rate of each frame could be calculated as follows: 1) The allocated bts of the vrtual GOP s frst calculated btsvgop = tran _ rate* wdth * heght *1.5* gop _ sze 1+ KeyframesIn1VGOP +.5* KeyframesIn VGOP),8) where the wdth and heght s the wdth and heght of encoded pctures respectvely, KeyframesInVGOP s the number of key frames ncluded by ths vrtual GOP and another neghbor vrtual GOP, and KeyframesIn1VGOP s the number of key frames that s ncluded just by ths vrtual GOP. ) Accordng to the schedulng sequence, for each FGS packet k n vrtual GOP f btspacketk < btsvgop ), the allocated bts of the packet k s frame s ncreased by the sze of packet k, and btsvgop = btspacket ; k else the allocated bts of the packet k s frame s ncreased by the remanng btsvgop. At last, the allocated bt rate of key pctures n two vrtual GOP s added together. 3 Comparson of the SASQR and JSVM_QL algorthms n the Foreman sequence JSVM_QL_.15 SASQR_.15 38 Comparson of the SASQR and JSVM_QL algorthms n the Moble sequence JSVM_QL_.15 SASQR_.15 PSNR [db] 1 39 38 PSNR [db] 36 3 3 37 3 36 5 1 15 5 3 35 GOP No. 8 5 1 15 5 3 35 GOP No. Fg. Comparson of JSVM-9 default qualty layer extracton and the proposed SASQR at the bt rate.15 bts/sample n the Foreman and Moble CIF sequences.. Expermental Results Smooth qualty reconstructon s the favorable applcaton. To valdate the effectveness of the SASQR, the SVC reference software JSVM 9, verson JSVM_9_) [8] s used. Eght standard sequences Foreman, Moble, Bus, Football, Cty, Crew, Harbour, and Soccer) are encoded at CIF resoluton. The SNR scalable confguraton fle of JVT-Q9 [9] s used wth four parameters changed: the GOP sze s set to 8, the FREXT mode s on for usng adaptve * or 8*8 transform, the number of FGS Levels s set to, the predcton loop s operated at the hghest FGS level. The target bt rate of FGS EL s.5 bts/sample about 8 kbts/s),.1 bts/sample about 56 kbts/s) and.15 bts/sample about 68 kbts/s) respectvely n the three experments. The smoothed GOPs are from to 36 n the Foreman, Moble, Cty,
shows Crew, Harbour, and Soccer sequences, from to 31 n the Football sequence, and from to 17 n the Bus sequence. Two dfferent algorthms are appled n the experments. The frst s the proposed SASQR. The default qualty layer method s also appled for comparson. Table. Comparson of JSVM-9 Default Qualty Layer Extracton and the Proposed SASQR n the Eght Standard CIF Sequences Sequences Foreman Moble Football Bus Cty Crew Harbour Soccer JSVM-9 QLdB) Proposed SASQRdB) Exp1 Exp Exp3 Exp1 Exp Exp3 Mn 3.88 36.58 36.71 36.9 38.55 39.66 Max 39. 1.66.7 38.33 39.96.69 Var 1.1.3.3.1.1.8 Mn 9.61 31.67 3.6 3.15 3.39 33.3 Max 33.5 33.8 37.15 3.1 33.39 3.33 Var.1..7.13..5 Mn 33.3 33.97 35.37 35.3 36.1 37.8 Max 3.85.77.77 39.65 39.98.31 Var 7.66 6.1.19.71.. Mn 3.1 3.98 3.98 33.36 3.85 35.91 Max 3.88 37.16 38.76 3.37 35.63 36.7 Var.55.33 1.1.8.3. Mn 35.76 39.8 39.13 37.55 39.7. Max 39.31 1.77 1.8 39..1 1.36 Var 1.1.71.63.11.5.5 Mn 3.9 36.78 37.1 36.51 37.9 38.76 Max.9 1.1.8 37.8 38.8 39.87 Var.1 1..3.8.6.6 Mn 3.83 3.98 35.51 33.31 3.69 35.75 Max 35.8 35.8 36.56 3.39 35.57 36.5 Var.99.7.15.6.1.3 Mn 3.9 36.99 38.8 36.5 38.6 39. Max.9.17.19 38.7 39.57.65 Var 3.8 1.51 3.5.1.7.6 HFg. H compares the proposed SASQR wth JSVM-9 default qualty layer extracton at the bt rate.15 bts/sample n the Foreman and Moble sequences, respectvely. HTableH shows the correspondng varaton statstcs of all the three experments n the eght test sequences. It can be seen that, compared to the JSVM-9 default qualty layer extracton, the smoother qualty can be obtaned wth the proposed SASQR. HFg. 5H the comparson of subjectve qualty between JSVM-9 default qualty layer extracton and the proposed SASQR at the bt rate.15 bts/sample n the frames 6 and 6 of the Foreman sequence. It can be seen that the subject dfference of frame 6 between JSVM-9 default qualty layer extracton and the proposed SASQR s lttle though the correspondng PSNR of JSVM-9 default method s.7 db hgher than the one of the proposed SASQR. However the subjectve qualty of frame 6 s mproved notceably n the proposed SASQR and the overall subjectve qualty s smoothed and mproved. 5. CONCLUSION In ths paper, frst, the vrtual GOP concept s ntroduced and a vrtual-gop-based prorty settng algorthm s proposed to obtan the optmal prorty order of FGS levels n a vrtual GOP. Second, a new effcent and flexble R-D model s presented for approxmatng the R-D functon of vrtual GOPs, wth whch a smple algorthm for smooth qualty reconstructon s employed to mprove the subjectve qualty of the reconstructed vdeo. Extensve experments show the effectveness and effcency of the models and algorthms.
a) JSVM-9 default qualty layer extracton b) The proposed SASQR Fg. 5. Subject comparson of JSVM-9 default qualty layer extracton a) and the proposed SASQR b) at the bt rate.15 bts/sample n the frame 6 and 6 of the Foreman CIF sequence. REFERENCES 1. H. Schwarz, D. Marpe, and T. Wegand, Overvew of the scalable vdeo codng extenson of the H.6/AVC standard, IEEE Trans. Crcuts. Syst. Vdeo Technol., vol. 17, no. 9, pp. 113 11, Sep. 7.. J. Rdge and M. Karczewcz, AHG report: FGS applcatons and desgn smplfcaton, Jont Vdeo Team JVT) of ISO-IEC MPEG & ITU-T VCEG, JVT-X6 Jul. 7. 3. I. Amonou, N. Cammas, S. Kervadec, S. Pateux, Optmzed rate dstorton extracton wth qualty layers n the scalable extenson of H.6/AVC, IEEE Trans. Crcuts. Syst. Vdeo Technol., vol. 17, no. 9, pp. 1186 1193, Sep. 7.. T. V. Lakshman, A. Ortega, and A. R. Rebman, VBR vdeo: Tradeoffs and potentals, n Proc. IEEE, vol. 86, May 1998, pp. 95 97. 5. J. Sun, W. Gao, D. Zhao, Statstcal analyss and modelng of rate-dstorton functon n SVC fne-granular SNR scalable vdeos, IEEE Internatonal Conference on Multmeda. & Expo, ICME 7, Bejng, Jul. 7. 6. J. Sun, W. Gao, D. Zhao, Rate-dstorton modelng and t s applcaton to qualty layer assgnment n SVC fnegranular SNR scalable vdeos, Pcture Codng Symposum, PCS 7, Lsbon, Nov. 7. 7. J. Sun, W. Gao, D. Zhao, and Q. Huang, Statstcal model, analyss and approxmaton of rate-dstorton functon n MPEG- FGS vdeos, IEEE Trans. Crcuts Syst. Vdeo Technol., Vol. 16, No., pp. 535-539, Apr. 6. 8. J. Rechel, H. Schwarz, M. Wen, and J. Veron, Jont Scalable Vdeo Model 9 of ISO/IEC 196-1:5/AMD3 Scalable Vdeo Codng, Jont Vdeo Team JVT) of ISO-IEC MPEG & ITU-T VCEG, JVT-X. Geneva, Jul. 7. 9. M. Wen, and H. Schwarz, AHG on Codng eff & JSVM codng effcency testng condtons, n Jont Vdeo Team JVT), Docs. JVT-Q9, Oct. 5.