RealVideo 10 "Winter 2004" update

[karl_lillevold]

** 2004-12-06: Fix for rcBestPSNRMode introduced 11-24.

rezard found a problem in which I accidentally always enabled a feature that was supposed to only be used with rcBestPSNRMode, so please update when you have a chance.

[Sirber]

RealAnime 2.30 will be ready with it.

[rezard]

the Rate Control stretch tiny frames.

RealVideo Rate Control 2nd pass log:
Framerate: 23976/1000 = 23.976000 fps
Total number of frames: 30518
Target filesize: 123833344 bytes
Keyframe boost adj: 1.00 Nonlinear adj: 1.00
CtrlStrngth: 5 MaxOFlowImpr: 1 MaxOFlowDegr: 1
Num Type Quant Size Scaled AdvSc Target Actual Error OFlow
0 i 14 19333 7627 7627 7627 9216 1589 -1589.00
1 p 13 5836 2581 2581 2555 2313 -242 -1387.74
2 p 13 8262 3746 3746 3708 2812 -896 -532.49
3 p 13 7850 3611 3611 3585 2645 -940 366.77
4 p 14 11903 4978 4978 4995 4364 -631 957.03
5 p 14 10873 4460 4460 4504 4059 -445 1361.28
6 p 13 8414 3821 3821 3859 3600 -259 1579.54
7 p 13 8376 3788 3788 3825 3127 -698 2236.79
8 p 13 11388 4773 4773 4820 4767 -53 2249.05
9 p 13 10531 4315 4315 4358 4449 91 2117.31
...
3297 p 5 892 459 459 1018 3002 1984 1152267.81
3298 p 3 453 240 240 798 3267 2469 1149796.56
3299 p 5 1083 563 563 1120 2639 1519 1148275.31
3300 p 7 1789 907 907 1463 2487 1024 1147249.06
3301 p 7 1517 808 808 1363 2211 848 1146398.80
3302 p 6 1367 714 714 1269 2526 1257 1145139.55
3303 p 7 1430 749 749 1303 2153 850 1144287.30
3304 p 5 960 499 499 1053 2948 1895 1142390.05
3305 p 5 918 468 468 1021 2614 1593 1140794.79
3306 p 4 725 378 378 930 2834 1904 1138888.54
3307 p 6 1101 576 576 1127 2180 1053 1137833.29
3308 p 5 1066 571 571 1122 2744 1622 1136209.04
3309 p 7 1622 819 819 1369 2389 1020 1135186.79
3310 p 5 806 425 425 975 2854 1879 1133305.53
3311 p 7 1559 817 817 1366 2250 884 1132419.28
3312 p 8 2105 1010 1010 1558 2397 839 1131578.03
3313 p 8 1886 900 900 1448 2222 774 1130801.78
3314 p 7 1337 677 677 1224 2241 1017 1129782.52
3315 p 8 1967 950 950 1497 2247 750 1129030.27
3316 p 8 1776 906 906 1453 2041 588 1128440.02
3317 p 9 2051 959 959 1505 2151 646 1127791.77
3318 p 9 2912 1415 1415 1961 2343 382 1127407.51
3319 p 8 2062 1011 1011 1557 2255 698 1126707.26
3320 p 9 2232 1117 1117 1663 2034 371 1126334.01
3321 p 8 1915 964 964 1509 2080 571 1125760.76
3322 p 9 2094 1041 1041 1586 1967 381 1125377.50
3323 p 8 1501 710 710 1255 2132 877 1124498.25
3324 p 8 1561 718 718 1262 2057 795 1123701.00
3325 i 15 29926 11020 11020 11130 13880 2750 1120951.00
3326 p 10 24994 9416 9416 9959 16862 6903 1113993.00
3327 p 12 11881 4740 4740 5279 4956 -323 1114261.00
3328 p 13 15048 6191 6191 6730 4956 -1774 1115980.00
3329 p 13 20556 8106 8106 8646 9054 408 1115517.00
3330 p 13 11945 4955 4955 5495 4445 -1050 1116512.00
3331 p 13 10167 4170 4170 4711 4034 -677 1117134.00
3332 p 14 23325 8987 8987 9528 9160 -368 1117447.00
3333 p 13 12313 4882 4882 5423 5275 -148 1117540.00
3334 p 13 11753 4891 4891 5432 4348 -1084 1118569.00
...

[Sagittaire]

CtrlStrngth: 5 MaxOFlowImpr: 1 MaxOFlowDegr: 1

you must use higher Oflow because quant prediction isn't very good

CtrlStrngth: 5 MaxOFlowImpr: 50 MaxOFlowDegr: 50

But for me the RV10 RC isn't good: for me a good RC could make second pass with constant quantizer (for I,P,B with possible ratio/offset for I quant boost and B quant reduction) with RC variability setting ...

XviD or RV10 use rcLowBitrateBoost and rcHighBitrateReduce for variability setting but I don't understand the difference between these setting: if you reduce bitrate for "high motion" (high bitrate part) you boost automaticaly bitrate for "low motion" (low bitrate part). Only one setting is necessary: varaibility like for x264 or Bitrate Modulation for DivX ...

For me the best example is x264 RC ...

[rezard]

I tried 5/50/50.
indeed, the number of extreme low quants were much less. but not perfect.

RealVideo Rate Control 2nd pass log:
Framerate: 23976/1000 = 23.976000 fps
Total number of frames: 30518
Target filesize: 123833344 bytes
Keyframe boost adj: 1.00 Nonlinear adj: 1.00
CtrlStrngth: 5 MaxOFlowImpr: 50 MaxOFlowDegr: 50
Num Type Quant Size Scaled AdvSc Target Actual Error OFlow
0 i 14 19333 7627 7627 7627 9216 1589 -1589.00
1 p 13 5836 2581 2581 2503 2261 -242 -1387.74
2 p 13 8262 3746 3746 3678 2782 -896 -532.49
3 p 13 7850 3611 3611 3585 2645 -940 366.77
4 p 14 11903 4978 4978 4995 4364 -631 957.03
5 p 14 10873 4460 4460 4506 4061 -445 1361.28
6 p 13 8414 3821 3821 3886 3627 -259 1579.54
7 p 13 8376 3788 3788 3864 3166 -698 2236.79
8 p 13 11388 4773 4773 4881 4828 -53 2249.05
9 p 13 10531 4315 4315 4423 4514 91 2117.31
...
6285 p 10 8181 3728 3728 4921 4944 23 24583.59
6286 p 11 7971 3517 3517 4708 4351 -357 24910.06
6287 p 11 10966 4918 4918 6125 5655 -470 25349.53
6288 p 11 7709 3521 3521 4749 4005 -744 26063.00
6289 i 7 2656 2425 2425 2465 2389 -76 26139.00
6290 p 9 2086 1168 1168 1801 1640 -161 26300.54
6291 p 7 954 565 565 1202 1831 629 25672.07
6292 p 5 467 268 268 890 2326 1436 24236.61
6293 p 3 184 161 161 748 2302 1554 22683.14
6294 p 1 465 256 256 805 2411 1606 21077.68
6295 p 2 482 247 247 757 2351 1594 19484.21
6296 p 4 862 436 436 908 2433 1525 17959.75
6297 p 6 747 356 356 791 1804 1013 16947.28
6298 p 7 931 427 427 837 1859 1022 15925.82
6299 p 9 1657 775 775 1162 1906 744 15182.35
6300 p 9 1650 735 735 1102 1939 837 14345.89
6301 p 10 1719 768 768 1152 1776 624 13722.42
6302 p 10 1635 724 724 1086 1769 683 13039.96
...

can I request you to explain about the MaxOFlowImpr/MaxOFlowDegr?

I'm eager to have a quant-limit option in the RC.

[Dark-Cracker]

@rezard
do u think u can try the RC AltCC who is in autorv10 to see which result it give u ?

++

[rezard]

sorry, I love manual encoding.

what should I do with .pass(analysis) file produced by ScaleCurve.exe?

[karl_lillevold]

I have on my list of things to do, to do another pass at the rate control. These are good suggestions for areas to look into:
1) QP limits.
2) improve prediction of QP for small frames.
3) perhaps combine rcLowBitrateBoost and rcHighBitrateReduce

Thanks.

[Sagittaire]

@ karl_lillevold

I test seriously the RV10 Rate Control ... and I humbly make some suggestion to improve it.


1) Don't use Average PSNR for RC optimization. Overall PSNR is very better for visual optimisation tweak. Use Overall PSNR in rv9log.txt.


2) Overall PSNR is better with quality mode (vbrQuality + scalingFactor). The quant prediction for New RC is not good for me. If second pass average quant is very higher than the first pass rcPFrameRefQuant then variability quant in 2 pass could be very high. That observation is confirmed by the high default overflow setting. Perhabs multipass rate control could be very better. First pass for first evaluation, second pass with first pass average quant prediction and final pass with better quant prediction. Perhabs bframe sensivity/treshold setting could be good too for test ...


3) bframe decision seem strange: 1 or 3 bframe but not 2. Bframe decision in P(n-2) B (n-1) I(n) with Iframe scene change frame is strange too:


23160 17 1752 40.08 23280 34.11 41.26 12943 100
23200 17 1508 39.96 21640 34.11 41.26 12958 100
23240 17 1180 40.35 19096 34.11 41.26 12969 100
23280 23 2120 35.15 17800 34.11 41.24 13150 d 100 ????
23320 17 13696 37.50 113648 34.11 41.25 13142 Key 100
23360 23 1304 35.37 13128 34.11 41.23 13230 d 100
23400 17 4920 36.91 60272 34.11 41.24 13230 100


IMO if you change bframe for Pframe the quality for Pframe is very higher (Pframe use previous frame, bframe use previous frame and Iframe scene change for reference ... :eek: ). Min frame PSNR is praticaly always for bframe in these particular sequences. IMO with Iframe scene change you must close GOP ...

[karl_lillevold]

Sagittaire: these are good suggestions. Thanks!

[dexx]

Can the RV10 codec be used with virtualdub for encoding?
Has there been any further updates since 10/2004?

[karl_lillevold]

no, and no. Sorry.

[stax76]

no, and no. Sorry.

Is there (COM+) automation or cli and a config dialog that can be invoked or any other effort, interface or API to support developers of batch encoding applications like Gordian Knot. Supporting a codec usually is very painful, XviD is the only exception.

[Sirber]

You can use helix producer commandline, like every other RV devs used.

[Castor-Troy]

@ karl_lillevold

I test seriously the RV10 Rate Control ... and I humbly make some suggestion to improve it.


1) Don't use Average PSNR for RC optimization. Overall PSNR is very better for visual optimisation tweak. Use Overall PSNR in rv9log.txt.


2) Overall PSNR is better with quality mode (vbrQuality + scalingFactor). The quant prediction for New RC is not good for me. If second pass average quant is very higher than the first pass rcPFrameRefQuant then variability quant in 2 pass could be very high. That observation is confirmed by the high default overflow setting. Perhabs multipass rate control could be very better. First pass for first evaluation, second pass with first pass average quant prediction and final pass with better quant prediction. Perhabs bframe sensivity/treshold setting could be good too for test ...


3) bframe decision seem strange: 1 or 3 bframe but not 2. Bframe decision in P(n-2) B (n-1) I(n) with Iframe scene change frame is strange too:


23160 17 1752 40.08 23280 34.11 41.26 12943 100
23200 17 1508 39.96 21640 34.11 41.26 12958 100
23240 17 1180 40.35 19096 34.11 41.26 12969 100
23280 23 2120 35.15 17800 34.11 41.24 13150 d 100 ????
23320 17 13696 37.50 113648 34.11 41.25 13142 Key 100
23360 23 1304 35.37 13128 34.11 41.23 13230 d 100
23400 17 4920 36.91 60272 34.11 41.24 13230 100


IMO if you change bframe for Pframe the quality for Pframe is very higher (Pframe use previous frame, bframe use previous frame and Iframe scene change for reference ... :eek: ). Min frame PSNR is praticaly always for bframe in these particular sequences. IMO with Iframe scene change you must close GOP ...

Sagittaire: these are good suggestions. Thanks!

No hope to see thoses tweaks :( ?