An efficient frame skipping transcoding from H.264/AVC to H.264/AVC including mode decision and motion vector decision methods had been proposed. In our propose methods, we obtain some information from the compressed video stream in H.264/AVC and then reuse them to decide the block mode types and motion vectors in the retained frame.
Our propose methods save more than 50% transcoding time and visual quality only reduced less than 0.2dB in most of the test sequences when comparing with H.264/AVC. Simulation results show that when comparing with all 16x16 mode, the propose methods improve the 0.2~0.3dB in PSNR measurement and reduce a lot of bit-rate. Besides, the experimental results also show that our propose methods improve about 0.1dB in average while comparing with no mode change. In future, we consider skipping not only one frame to make more decision about the block mode types and the length of motion vectors.
Reference
[1] Draft ITU-T Recommendation and Final Draft International Standard of Joint Video Specification (ITU-T Rec. H.264-ISO/IEC 14496-10AVC), JVT-G050, Joint Video Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, May 2003.
[2] G. Keesman, R. Hellinghuizen, F. Hocksema, and G. Heideman, “Transcoding of MPEG-2 bitstreams,” Signal Processing: Image Comm., vol. 9, pp. 481-500, 1996.
[3] N. Bjork and C. Chistopolous, “Transcoder architecture for video coding,” IEEE Trans. Consumer Electron., vol. 44, pp. 88-98, Feb. 1998.
[4] Chia-Wen Lin, Yuh-Reuy Lee, “Fast algorithms for DCT-domain video transcoding,” Proc. IEEE Int. Conf. Image Processing, vol. 1, pp. 421-424, Oct.
2001.
[5] Kai-Tat Fung, Yui-Lam Chan, Wan-Chi Siu, “A new architecture for dynamic frame skipping transcoder,” IEEE Trans. On Image Processing, vol. 11, pp.
886-900, Aug. 2002.
[6] Mei-Juan Chen, Ming-Chung Chu, Chih-Wei Pan, “Efficient motion-estimation algorithm for reduced frame-rate video transcoder,” IEEE Trans. Circuits Syst.
Video Technol., vol. 12, pp. 269-275, Apr. 2002.
[7] J. Youn, M. –T. Sun, and C. –W. Lin, “Motion vector refinement for high performance transcoding,” IEEE Trans. Multimedia, vol. 1, pp. 30-40, Mar.
1999.
[8] Yusuf A. A., Murshed M. and Dooley L. S., “An Adaptive Motion Vector Composition Algorithm for Frame Skipping Video Transcoding,” IEEE Electrotechnical Conference, vol. 1, pp. 235-238, May. 2004.
[9] Joint Video Team, “Reference software JM13.2,”
http://iphome.hhi.de/seuhring/tmldowload.
[10] Feng Pan, Z. P. Lin, and X. Lin, “Content adaptive frame skipping for low bit rate video coding,” ICICS-PCM, pp. 15-18, December 2003.
[11] F. Lonetti and F. Martelli, “Motion vector composition algorithm in H.264 transcoding,” 6th EURASIP Conference focused on Speech and Image Processing. 14th International Workshop, pp. 401-404, June 2007.
[12] J.N. Hwang, T.D. Wu, C.W. Lin, “Dynamic frame-skipping in video transcoding,” Proceedings of 2nd Workshop on Multimedia Signal Processing, pp. 616-621d, December 1998.
Appendix A
Block Mode Type Observation
In this section, we observed about the block mode type relations between current frame and previous frame. Here, we abided by the relative macroblock location set up by the JM reference software—X, A, and B, as Fig. A.1 shows:
X A
B
Previous Frame Current Frame
Encoded MB
Unencoded MB
Current MB
Fig. A.1 Macroblock relative location: X, A, and B
where X, A, and B represent the co-located, left, and up macroblock in the previous frame, respectively. Here we conducted different CIF sequences with 100 frame numbers per test sequence as inputs. The test sequences are Foreman, Football, Tennis, Stefan, Bus, Container, and Hall. Each kind of macroblock type, PSKIP, P16x16, P16x8, P8x16, SMB8x8, SMB8x4, SMB4x8, SMB4x4, Intra16x16, and Intra4x4, would be checked. Fig. A.2 exhibits each test sequence with 100 frames that contains the proportion of different kinds of macroblock type.
0 %
Fig. A.2 Statistic different MB types of each 100 test sequences in percentage (%)
Here, we divided different kinds of macroblock types into three parts. The first part is large macroblock type which includes PSKIP, P16x16, P16x8, and P8x16. The second part is called small macroblock type which contains from SMB8x8, SMB8x4, SMB4x8, and SMB4x4. The final part is intra macroblock type that comprises Intra4x4 and Intra16x16. The statistical data and the bar chart are shown as following
in Table A.1 and figure A.3.
se q ue nc e Large Small intra fo o t b a ll 25.65% 70.87% 3.47%
st e fa n 33.36% 55.24% 0.72%
b u s 37.22% 62.31% 0.47%
te nn is 44.48% 53.30% 2.22%
fo re m a n 61.25% 37.38% 1.37%
h a ll 76.83% 21.76% 1.41%
co n ta in e r 83.36% 16.17% 0.47%
Table A.1 Catalog MB type in percentage (Large: PSKIP~P8x16, Small: SMB8x8~SMB4x4)
0%
1 0%
2 0%
3 0%
4 0%
5 0%
6 0%
7 0%
8 0%
9 0%
1 0 0%
fo o tb a ll st e fa n b u s t e n n is fo re m a n h a ll co n t a in e r
Se q u e n ce
Number La rg e
Sm a ll in t ra
Fig. A.3 Catalog MB type in percentage (Large: PSKIP~P8x16, Small: SMB8x8~SMB4x4)
Since the information mentioned above, we could conclude that higher percentage of larger macroblock obtained from test sequence means the test sequence is close to slow motion one. On the contrary, if the test sequence has higher percentage of small macroblock, it would be classified into high or fast motion sequence.
Table A.2 and Table A.3 illustrate the most and second frequent macroblock type appearance would be taken out and the statistical data would be revealed. Table A.1
shows the situation when current macroblock type is PSKIP.
se q ue nc e PSKIP P16x16 P16x8 P8x16 SMB8x8 fo re m a n 11777 12935 3773 4880 13132 fo o t b a ll 7640 6919 3062 3849 29573 t e nn is 10899 8629 4535 4987 20691
st e fa n 8422 7587 3068 2715 18968
b u s 6609 10796 5200 4787 26828
co n t a in e r 28960 5096 1609 1527 5079
h a ll 24064 8120 2515 1295 6349
se q ue nc e SMB8x4 SMB4x8 SMB4x4 I16MB I4MB
fo re m a n 3043 3392 797 350 398
fo o t b a ll 11454 14483 3802 34 2872
t e nn is 6236 6374 1515 1028 422
ste fa n 6963 7265 2884 283 189
b u s 8520 8159 2349 79 269
co n t a in e r 823 969 345 205 3
h a ll 1523 1693 631 559 102
Table A.2 Statistic different MB types of each 100 test sequences
se q ue nc e PSKIP P16x16 P16x8 P8x16 SMB8x8 fo re m a n 21.92% 24.07% 7.02% 9.08% 24.44%
fo o t b a ll 9.46% 8.57% 3.79% 4.76% 36.61%
t e nn is 17.07% 13.51% 7.10% 7.81% 32.40%
st e fa n 13.19% 11.88% 4.80% 4.25% 29.70%
b u s 9.02% 14.74% 7.10% 6.54% 36.63%
co n t a in e r 65.21% 11.48% 3.62% 3.44% 11.44%
h a ll 52.10% 17.58% 5.44% 2.80% 13.75%
se q ue nc e SMB8x4 SMB4x8 SMB4x4 I16MB I4MB fo re m a n 5.66% 6.31% 1.48% 0.65% 0.74%
fo o t b a ll 14.18% 17.93% 4.71% 0.04% 3.56%
t e nn is 9.76% 9.98% 2.37% 1.61% 0.66%
st e fa n 10.90% 11.38% 4.52% 0.44% 0.30%
b u s 11.63% 11.14% 3.21% 0.11% 0.37%
co n t a in e r 1.85% 2.18% 0.78% 0.46% 0.01%
h a ll 3.30% 3.67% 1.37% 1.21% 0.22%
Table 5.3 Statistic different MB type in percentage
Appendix B
Motion Vector Observation
Observing and understanding the quantity of motion vector between different macroblock type and different test sequences is the goal of this experiment. In the experiments, we also examined different sequences including Foreman, Football, Tennis, Stefan, Bus, Container, and Hall. Table B.1 shows all kinds of inter macroblock with the rates of motion vector length of the sequence “Foreman”.
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up P16x16 22.72% 32.57% 20.64% 13.19% 7.75% 2.52% 0.55% 0.06%
P16x8 20.12% 27.46% 21.20% 14.55% 9.28% 5.17% 1.72% 0.50%
P8x16 23.57% 29.16% 22.42% 13.36% 8.03% 2.77% 0.57% 0.12%
SMB8x8 25.02% 28.65% 21.37% 12.79% 7.66% 3.72% 0.57% 0.22%
SMB8x4 24.88% 28.06% 20.97% 12.42% 8.61% 3.84% 0.95% 0.26%
SMB4x8 27.00% 27.86% 21.55% 12.85% 7.40% 2.62% 0.53% 0.18%
SMB4x4 28.86% 29.99% 23.71% 10.79% 4.14% 2.13% 0.25% 0.13%
Table B.1 All MB mode and its MV length in percentage (Foreman)
Notice th arter pixel in
H.264/A
acroblock with the rates of motion vector length of
at since the precision of motion vector can up to one-qu VC, the unit of the MV has multiplied by four.
Table B.2 to Table B.7 shows all kinds of inter m
the sequence Football, Tennis, Stefan, Bus, Container, and Hall.
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up P16x16 44.33% 36.84% 7.72% 3.44% 4.83% 2.02% 0.79% 0.03%
P16x8 34.16% 26.62% 9.76% 8.88% 10.32% 5.78% 3.40% 1.08%
P8x16 35.46% 28.58% 10.73% 7.01% 9.43% 5.07% 2.88% 0.83%
SMB8x8 31.26% 25.78% 12.33% 10.82% 10.65% 5.29% 2.95% 0.92%
SMB8x4 27.81% 24.20% 14.05% 12.79% 11.53% 5.49% 2.97% 1.16%
SMB4x8 27.01% 27.06% 14.46% 12.31% 10.96% 4.86% 2.58% 0.76%
SMB4x4 28.46% 29.17% 15.97% 11.65% 8.44% 3.76% 1.76% 0.79%
Table B.2 All MB mode and its MV length in percentage (Football)
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up P16x16 28.15% 17.52% 16.93% 24.28% 8.61% 3.13% 1.08% 0.30%
P16x8 25.47% 16.67% 17.95% 22.27% 10.32% 4.61% 2.14% 0.57%
P8x16 22.80% 24.38% 19.71% 21.05% 7.98% 2.47% 1.32% 0.28%
SMB8x8 29.33% 21.60% 17.35% 17.97% 8.38% 3.51% 1.47% 0.40%
SMB8x4 31.03% 20.33% 16.31% 16.95% 8.32% 4.39% 2.05% 0.61%
SMB4x8 26.80% 22.11% 18.28% 18.34% 8.53% 4.31% 1.30% 0.33%
SMB4x4 34.65% 20.99% 13.53% 16.57% 8.38% 3.50% 1.91% 0.46%
Table B.3 All MB mode and its MV length in percentage (Tennis)
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up P16x16 19.20% 12.64% 7.75% 13.52% 26.52% 10.18% 8.61% 1.58%
P16x8 23.04% 10.37% 7.95% 13.43% 24.25% 10.07% 8.60% 2.28%
P8x16 19.56% 10.17% 8.88% 14.73% 25.34% 10.17% 9.54% 1.62%
SMB8x8 19.56% 10.17% 8.88% 14.73% 25.34% 10.17% 9.54% 1.62%
SMB8x4 16.37% 10.21% 7.77% 15.11% 27.03% 11.03% 9.72% 2.76%
SMB4x8 14.87% 9.51% 9.64% 16.30% 29.68% 9.69% 8.51% 1.82%
SMB4x4 17.16% 9.15% 8.04% 15.71% 27.88% 11.10% 9.29% 1.66%
Table B.4 All MB mode and its MV length in percentage (Stefan)
block\mv len. 0 100 101up
P16x16 2.92% 2.52% 5.05% 12.04% 50.03% 14.17% 10.85% 2.43%
P16x8 4.00% 4.63% 8.12% 13.81% 41.38% 14.15% 11.00% 2.90%
P8x16 3.01% 2.88% 5.49% 11.13% 44.52% 16.61% 14.10% 2.26%
SMB8x8 6.31% 5.36% 8.78% 12.27% 42.08% 12.80% 10.52% 1.87%
SMB8x4 12.17% 5.70% 9.98% 12.66% 37.59% 10.86% 9.32% 1.71%
SMB4x8 6.73% 6.73% 10.27% 12.92% 39.51% 11.31% 10.81% 1.72%
SMB4x4 9.92% 6.60% 11.92% 11.20% 38.78% 10.34% 9.75% 1.49%
~1 2~4 5~8 9~15 16~30 31~50 51~
Table B.5 All MB mode and its MV length in percentage (Bus)
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up
P16x16 55. 0.22% 0.12%
P16x8 57.18% 38.66% 1.43% 0.56% 0.68% 0.37% 0.75% 0.37%
P8x16 53.57% 41.52% 2.16% 0.52% 1.18% 0.39% 0.39% 0.26%
SMB8x8 67.18% 30.48% 1.40% 0.32% 0.47% 0.00% 0.16% 0.00%
SMB8x4 71.45% 25.76% 1.70% 0.12% 0.97% 0.00% 0.00% 0.00%
SMB4x8 69.25% 29.41% 1.14% 0.00% 0.21% 0.00% 0.00% 0.00%
SMB4x4 64.06% 32.17% 2.32% 0.87% 0.58% 0.00% 0.00% 0.00%
87% 40.70% 1.31% 0.67% 0.96% 0.16%
Table B.6 All MB mode and its MV length in percentage (Container)
block\mv len. 0~1 2~4 5~8 9~15 16~30 31~50 51~100 101up P16x16 93.74% 2.54% 1.15% 1.42% 0.85% 0.25% 0.05% 0.01%
P16x8 92.41% 2.82% 1.43% 1.91% 1.11% 0.20% 0.12% 0.00%
P8x16 87.49% 3.86% 3.09% 3.01% 1.93% 0.54% 0.08% 0.00%
SMB8x8 79.79% 7.51% 4.90% 4.55% 2.35% 0.60% 0.25% 0.05%
SMB8x4 71.63% 11.29% 6.83% 6.37% 3.41% 0.07% 0.33% 0.07%
SMB4x8 59.30% 14.83% 10.63% 8.56% 4.61% 0.95% 1.12% 0.00%
SMB4x4 68.30% 13.95% 7.77% 6.50% 2.06% 0.79% 0.63% 0.00%
Table B.7 All MB mode and its MV length in percentage (Hall)
ppendix C
DVS, ADVS, and PADVS(n) Methods on Large Block Size
Originally, all of the methods such as FDVS, ADVS, and PADVS(n) are used on PEG-2 video standard. Unlike H.264/AVC, the macroblock modes of MPEG-2 are 2 video standard contains only four modes, SKIP otion compensation, and inter mode with zero motion. For the sake of applying to our variable block size in H.264/AVC, we simulate the operation by closing some mode search in H.264/AVC. Hence, we close the following mode search parameters in reference software JM:
PSliceSearch16x8 = 0 PSliceSearch8x16 = 0 PSliceSearch8x8 = 0 PSliceSearch8x4 = 0 PSliceSearch4x8 = 0 PSliceSearch4x4 = 0
And then we reserve the mode search of P16x16, PSKIP, and intra16x16 which are all size of 16x16 block modes.
The meaning of FDVS method is to find out the dominant macroblock that has the largest overlapping area pointed by the current motion vector. However, due to the motion vector precision up to quarter-pixel in H.264/AVC, calculating the overlapping area is not a nice manner since it is a little redundant. So, we propose a better way to find out largest overlapping area instead. Figure C.1 illustrates the FDVS method on all 16x16 modes.
A
F
M
not as many as H.264/AVC. MPEG-mode, intra MPEG-mode, inter mode with m
frame (n-1) Current MV1 Current MV2
frame (n) 16
16 64 in quarter
64 in quarter
Fig. C.1 Illustrates the FDVS method on all 16x16 modes
There exist two current motion vectors, MV1 and MV2. The blue macroblock is the target one that we want to find its dominant motion vector in frame (n). Suppose we have two different quantities of current motion vectors as following,
And then we separate the motion vector by x and y direction. As shown in figure C.2.
frame (n) frame (n) Current MV1 x-dir
y-dir
Current MV2 x-dir
y-dir
Fig. C.2 Separate the current MV in x and y direction
After separating the motion vector in the coordination of x and y, we use the following formula to find out the dominant macroblock.
32
Qx and Qy are quantities of the shift offset location in x and y direction. In out examples, current MV1 would occupy the blue macroblock which is the dominant macroblock while current MV2 would occupy the purple macroblock as dominant one.
We can find out the dominant motion vector from the following array then.
Finally, we finish finding out the dominant motion vector by using FDVS method on all 16x16 block modes.
Like FDVS, ADVS method is also used in MPEG-2 video standard. To analyze the ADVS method, we nose out that ADVS is much complex than FDVS because we have
C.3 and C.4 illustrate situations that the ADVS meth
to calculate the number of non-zero quantized coefficients that the area is covered by the macroblock. Figure
od covered on the 8x8 block.
frame (n-1) 8x8 non-zero frame (n-1)
quantized coefficients
Fig. C.3 The first case of ADVS covering on all 16x16 block modes
In the first case shown in figure C.3, regardless of the quantities of current motion vector, the red dotted line represents the area of final result after adding current motion vector. What we care about is the direction of x and y, that is we want
to know wheth t case, except
the c
Observe the
er they are positive or negative. The upper one in the firs
enter macroblock covered four 8x8 sub-blocks, we also need to calculate two 8x8 sub-blocks above and next to the center macroblock. Finally, still one 8x8 sub-block at the corner of the remaining covered macroblock is needed to be counted. So do the same operations in figure C.4.
frame (n-1) frame (n-1) Observe the
8x8 non-zero quantized coefficients
Fig. C.4 The second case of ADVS covering on all 16x16 block modes
We sum up the rules of ADVS as the followings:
If the location of the center macroblock is called (x, y), see in figure C.5.
(x,y+1)
Fig. C.5 The representation of the ADVS location
Algorithm:
If x is positive, then calculate the right part of two 8x8 sub-blocks
then calculate the upper part at (x-1, y)
If y is positive, of two 8x8 sub-blocks at (x, y+1)
and upper-right part of one 8x8 sub-block at (x-1, y+1) negative, then calculate the lower part
If y is of two 8x8 sub-blocks at (x, y-1)
and lower-right part of one 8x8 sub-block at (x-1, y-1) If x is negative, then calculate the left part of two 8x8 sub-blocks at (x+1, y)
If y is positive, then calculate the upper part of two 8x8 sub-blocks at (x, y+1) and upper-left part of one 8x8 sub-block at (x+1, y+1)
If y is negative, then calculate the lower part of two 8x8 sub-blocks at (x, y-1) and lower-left part of one 8x8 sub-block at (x+1, y-1)
PADVS(n) method on all 16x16 block is similar with ADVS method. The only different between PADVS(n) and ADVS is the parameter n which n represents low frequency DCT coefficients in the estimation of activity. Besides, n is a set of DCT coefficients in the zigzag scan order. The set of is defined as follows,
}
the opinion of PADVS(n) is that using all of the 64 DCT coefficients to select the dominant macr
thus reducing the overall comp
ethods on large block size.
Since a 8x8 DCT quantized coefficients has 64 coefficients,
oblock is a computationally expensive process. Study of the human visual system (HVS) reveal that high frequency DCT coefficients usually have little impact on the perception except showing finer details. Although activity estimation is seemingly unrelated to image perception, the inherent correlation leads to the obvious query whether the impact of the high frequency DCT coefficients is minimal compared to that by the low frequency components, including the DC.
An affirmative response to this query would lead to ignore the high frequency coefficients in selecting the dominant macroblock and
utational complexity. This is an added benefit of ignoring the high frequency coefficients.
Table C.1 to Table C.4 are the experimental results of FDVS, ADVS, and PADVS(n) m
38.24 38.25 38.24 38.25 38.25
-0.34 -0.33 -0.34 -0.33 -0.33
Decoding tim 0 52.37 1 35.153 35.026 33.784 33.767 34.064 33.768 Encodin g tim 406.118 69.24 6 68.865 68.114 67.71 68.395 67.686 70.955
M.E. time 357.946 3.80 5 4.037 4.15 3.899 4.141 4.064 4.243
Total time 406.118 Transcoding
time 1 21.61 7 104.018 103.14 101.494 102.162 101.75 104.723
△T. time(%) 0 -70.05% -74.39% -74.6 0% -75.01% -7 4.84% -74.95% -7 4.21%
△M.E.time(% 0 -84.31% -89.05% -89.0 6% -89.47% -8 9.41% -89.35% -8 9.38%
Total bits 871368 98868 8 987720 994600 995000 994600 994000 994000 Bit-rate 217.84 247 .1 7 246.93 248.65 248.75 248.65 248.5 248.5
Table C.1 Experimental results of FDVS, ADVS and PADVS(n) on news.cif
n=1 n =3 n=10 n=36 n=49
PSNR 37.99 37 .8 5 37.85 37.85 37.85 37.85 37.85 37.85
△PSNR 0 -0 .1 4 -0.14 -0.14 -0.14 -0.14 -0.14 -0.14
Decoding tim 0 53.52 1 35.803 35.579 34.881 35.015 35.55 34.941 Encodin g tim 407.22 69.24 6 68.865 68.114 67.71 68.395 67.686 70.955
M.E. time 358.403 3.75 8 3.908 3.911 3.843 4.048 4.029 4.026
Total time 407.22 Transcoding
time 1 22.76 7 104.668 103.693 102.591 103.41 103.236 105.896
△T. time(%) 0 -69.85% -74.30% -74.5 4% -74.81% -7 4.61% -74.65% -7 4.00%
△M.E.time(% 0 -84.02% -88.92% -88.9 8% -89.20% -8 9.10% -88.96% -8 9.13%
Total bits 856472 99166 4 984448 993432 991904 991792 990880 991472 Bit-rate 214.12 247 .9 2 246.11 248.36 247.98 247.95 247.72 247.87
Item\Methods JM PADVS(n)
FDVS ADVS
Table C.2 Experimental results of FDVS, ADVS and PADVS(n) on hall.cif
n=1 n=3 n=10 n=36 n=4 9
PSNR 3 4.9 34 .7 4 3 4.69 34.69 34.7 34.7 34.7 34.69
△PSNR 0 -0 .1 6 -0.21 -0.21 -0.2 -0.2 -0.2 -0.21
Decod ing tim 0 71.36 6 50 .8 06 51.029 5 0.373 50.023 50.042 50.014 Encoding tim 422.698 152.03 1 149 .0 34 149.586 14 8.849 1 49.355 149.147 1 51.744 M.E. time 361.711 4.18 7 4 .1 48 4.306 4.247 4.249 4.139 4.247 Total time 422.698
Transcoding
time 223.39 7 19 9.84 200.615 19 9.222 1 99.378 199.189 2 01.758
△T. time(%) 0 -47.15% -52.72% -52.54% -52.87% -52.83% -52.88% -52.27%
△M.E.time(% 0 -79.11% -84.81% -84.70% -84.90% -85.00% -85.02% -85.00%
Total bits 5676504 579346 4 57932 80 5793280 5792 768 5794584 5797456 579 4504 Bit-rate 1419 .13 1448 .3 7 144 8.32 1448.32 14 48.19 1 448.65 1449.36 1 448.63
PADVS(n)
FDVS ADVS
Item\Methods JM
Table C.3 Experimental results of FDVS, ADVS and PADVS(n) on football.cif
n=1 n =3 n=10 n=36 n=49
PSNR 36.42 36 .3 5 36.31 36.31 36.31 36.31 36.31 36.31
△PSNR 0 -0 .0 7 -0.11 -0.11 -0.11 -0.11 -0.11 -0.11
Decoding tim 0 71.19 1 45.925 44.935 45.158 44.565 44.259 45.15 Encodin g tim 416.704 1 52.32 5 152.653 151.693 152.327 151.699 151.821 151.352
M.E. time 359.681 4.05 3 4.253 4.417 4.399 4.395 4.308 4.267
Total time 416.704 Transcoding
time 2 23.51 6 198.578 196.628 197.485 196.264 196.08 196.502
△T. time(%) 0 -46.36% -52.35% -52.8 1% -52.61% -5 2.90% -52.95% -5 2.84%
△M.E.time(% 0 -79.08% -86.05% -86.2 8% -86.22% -8 6.39% -86.50% -8 6.26%
Total bits 5257072 564687 2 5638584 5600928 5592240 5596264 5596264 5596264 Bit-rate 1313.27 1 411 .7 2 1409.65 1400.23 1398.06 1399.07 1399.07 1399.07
Item\Methods JM PADVS(n)
FDVS ADVS
Table C.4 Experimental results of FDVS, ADVS and PADVS(n) on flower.cif