Quant 2 for BFrames ?

[philippas]

I'm trying to set the quant for bframes = 2 but it doesn't work.
Quantizer ratio = 0 and Quantizer offset = 0 should result quant =2 (i think) but it doesn't and manuall range of quantizers 2-2 is ignored by the codec.
Is it harcoded in the codec to use a min bframe quant=3 ?

[Harley Quin]

Hi,
no, it's not hardcoded, at least not in the v1.10b2.
The quantizer is multiplied with the "Quantizer Ratio", then the "Quantizer Offset" is added. You did multiply with zero...
Use Ratio=1.00 and Offset=0.00 and you will get the same quantizer for all frames.
But keep in mind that the reduction in size, which comes with the use of b-frames, has its cause mostly due to the reduction of the quantizer, not the bi-directionality itself. Be sure to use "VHQ for b-frames" for b-frames to be of use at all. And compare it to an encoding without b-frames whether it is worth it.

[stephanV]

quant 2 for b-frames defeats the purpose of them

[Soulhunter]

But keep in mind that the reduction in size, which comes with the use of b-frames, has its cause mostly due to the reduction of the quantizer, not the bi-directionality itself...
Not true, at least if you only use 1 consecutive B-VOP...

With most stuff I tested B-VOPs were ~50% smaller than the P-VOPs !!!

http://img179.echo.cx/img179/8789/snap0289jo.png

Bye

[yaz]

Not true, at least if you only use 1 consecutive B-VOP... With most stuff I tested B-VOPs were ~50% smaller than the P-VOPs !!!confirmed ! this b-vop size issue is one of the most common misbeliefs.

imho, the default b-vop setting is an overdrive. setting it to 1/1/1 gave me the highest quality metrics in all cases(!) i've ever tested. besides, it produces a significant gain in compressibility too !

the bests
y

[Teegedeck]

Others have tested it several times with exactly the opposite result, so me thinks it depends on the source. But take it from someone who should know (sysKin): B-frames are meant to be higher compressed than p-frames. Full stop.

And, yaz, you use an offset of 1, so your b-frames are more strongly compressed than your p-frames.

[Soulhunter]

Others have tested it several times with exactly the opposite result...
Yeah, I also heard of this "rumor" !!!

But Ive never seen it so far... XD


Bye

[Teegedeck]

Well, I for one tested it; guess I've privately tested any and every feature. And I also found that b-frame-quant=p-frame-quant didn't yield any advantage over not using b-frames at all.

Guess I'm gonna have another go at it today. :p

Who knows, maybe b-vhq does the trick?

[stephanV]

I made a post about this more than a year ago: here (http://forum.doom9.org/showthread.php?t=76760&page=2&pp=20&highlight=b-frames)

To quote sysKin:

Mpeg-4 is much more optimized for low bitrates - so if difference between p-frames is bigger (due to bframes in between), it works worse, because p-frames were designed to carry very little data.
Also, b-frames are not expected to carry much data either, and are not optimized for that either (even more than p-frames). B-frames at the same quant are not what bitstream expects, so the compression is much worse.

Last but not least, p-frames are much better than in mpeg-2, while b-frames are very simplistic, almost like in mpeg-2.

But also:


It is also true that for xvid, p-frames have much better ME due to VHQ.

Maybe time for a small test again. :)

[Soulhunter]

1. Very grainy / Medium-motion / h.263

http://img300.echo.cx/img300/1903/moviegrainylomoh2633el.png


2. Some noise / Fast-motion / MPEG

http://img101.echo.cx/img101/4666/movieneutralhimompeg1ac.png


3. Neutral picture / Fast-motion / h.263

http://img210.echo.cx/img210/5390/movieneutralhimoh5fm.png


4. Very clean cartoon / Low-motion / h.263

http://img101.echo.cx/img101/2656/cartoonverycleanlomoh9zz.png


5. Grainy anime / High-motion / h.263

http://img101.echo.cx/img101/5108/animegrainyfamoh4ca.png

Bye

[Teegedeck]

First test: I still had my setup from testing a soccer clip, so I used that. Not your average source; DVB-T,strongly postprocessed. More tests to follow, so anyway:


Test 1: DVB-T soccer clip, postprocessed, denoised*
Filesizes:
No b-frames: 74.670 KB
Soulhunter's settings: 73.354 KB

So yes; your use of b-frames is superior to the use of no b-frames at all by a whopping 1.76% here.

*= Javor's 1-CD matrix @ quant=2, VHQ=4, B-VHQ, Trellis, AQ, GMC etc.

[stephanV]

Small test from me (1st minute of Sagittaire's HP2 trailer, using his script)

XviD-1.1.0beta2 (all settings default unless specified)


no bvops, VHQ4: 2755 kbps
bvops(1/1/0), VHQ4, no b-VHQ: 2727 kbps
bvops(1/1/0), VHQ4, b-VHQ: 2684 kbps
bvops(2/1/0), VHQ4, b-VHQ: 2690 kbps


max size diff of 2.6%

[Teegedeck]

2nd test: 20000 frames from Matrix Reloaded, half action, half talk. MPEG matrix @ quant=2

Filesizes:
No b-frames: 32.592 KB
Soulhunter's settings: 31.598 KB
XviD defaults: 23.218 KB

So here Soulhunter's b-frame settings save 3.04% over the use of no b-frames at all (thanks to b-vhq, as stephanV's results indicate). XviD's default settings save 28.76%, just for the record. ;)

Edit: BTW, Soulhunter, just looking at the size of individual b-frames doesn't help as a means of comparison - it is overall filesize differences that matter. A p-frame-only encode features overall smaller p-frames, than an encode with b-frames. When we have two p-frames, each 500 k, and on the other hand one p-frame of 750 k, followed by a b-frame of 250 k, the filesize difference between the two encodes still is naught.

[Soulhunter]

BTW, Soulhunter, just looking at the size of individual b-frames doesn't help as a means of comparison - it is overall filesize differences that matter. A p-frame-only encode features overall smaller p-frames, than an encode with b-frames. When we have two p-frames, each 500 k, and on the other hand one p-frame of 750 k, followed by a b-frame of 250 k, the filesize difference between the two encodes still is naught.

Yup, you are right! But throwing-in the size of XviDs defaults is also misleading, coz it tells you absolutely nothing about the efficiency of B-VOPs... Ok, the resulting filesize is 28.76% lower, but of coz the quality (visual n' metric wise) is lower as well! No matter, I stand correct, actually it seems to be very hard to produce a encode where B-VOPs are less efficient than P-VOPs... However, metric wise a Q2 B-VOP should be still worse than a Q2 P-VOP! So, at the end the over-all efficiency gain shrinks from 5-2% -> 3-1% or so...


Bye

[yaz]

Others have tested it several times with exactly the opposite result, so me thinks it depends on the source.nope. i've tested many different sources but haven't found any discrepancies. it's just a 'rumor', as soulhunter pointed out. even b110 provides some, mostly negligible, gain in compression.
But take it from someone who should know (sysKin): B-frames are meant to be higher compressed than p-frames. Full stop.when i'm testing i always go through a 000,110,111,121, ... sequence. and (believe or not) always (ok, in most instances) found 111 giving the highest quality metrics (w/significant gain in compression). i don't know what b-vop is meant for in general, and i don't really care, because i use a certain implementation from which i try to get the most ;)
And, yaz, you use an offset of 1, so your b-frames are more strongly compressed than your p-frames.actually, qb=qp+1. strong or not, i don't know.
however, i must add that i always try to maintain q avg round 3 (+-.5). i should have mentioned this too.

the bests
y

[Soulhunter]

...111 giving the highest quality metrics.

Same observation here... ^^ At least for constant-quality encodes it seems that 1/1/1 B-VOP settings give better results than the defaults. Unlike Q4 B-VOPs, Q3 B-VOPs look visually identical to Q2 P-VOPs for me! Tho, for low bitrates with higher quantizer fluctuation the defaults could/should still work better imo...


Bye

[Teegedeck]

However, metric wise a Q2 B-VOP should be still worse than a Q2 P-VOP! So, at the end the over-all efficiency gain shrinks from 5-2% -> 3-1% or so...Yes, you still have a small albeit considerable efficiency-gain of about 3% there but the merit of that for quality is perhaps questionable.

Turning from size-comparisons to the thing that realy interests us, the visual difference between my Q2 encodes with no b-frames, your b-frame settings and XviD's standard settings is not visible to me. I cannot tell the files apart on my 19'' Eizo Nanao TFT and I am quite a discerning viewer! Perhaps if I directly compared one still-frame against the other I could.*

But the decisive thing is, if I would encode that clip from Matrix Reloaded to 20 MB in two-pass, the clips with no b-frames and that with your b-frame quantizer settings would feature an average p-frame quant of around 3, while the average p-frame quant of the clip with standard settings would still be 2.

And I could see that difference very well. And so could anyone, even with less sophisticated equipment and a more casual glance. So for any purpose where filesizerestrictions of any kind exist (and this almost means 'always'), I would not recommend using ratio=1.00, offset=0.

Efficiency does translate into quality in the very end, and 25% is a heck of an efficiency discrepancy.

*=Edit: everyone his own. Anyone can make a blindtest and see how often out of 10 attempts one can tell an i/p/b-frame encode (default settings) from an i/p-frame encode.

[Teegedeck]

when i'm testing i always go through a 000,110,111,121, ... sequence. and (believe or not) always (ok, in most instances) found 111 giving the highest quality metrics (w/significant gain in compression). Those were comparisons at constant quantizer, it would seem. How about testing at identical filesize?

[Didée]

The technical analysis already was done by you mates, so let me throw in my "standard" (mostly) Bframe-settings:

2 - 1.63 - 0

Regarding quantizers, this effectively gives Bquant = Pquant + 1 for small quantizers (2 & 3), and for higher quants (q5 up to q10) is almost identical to XviD's defaults.

Using 6of9 pretty much, I'm mostly using a P-quantizer range of q3~q5/q6 (usually with some curve compression, if anyone still remembers this evil feature ;) ) - and in this scenario, I've settled for above settings.

[Teegedeck]

You always find a better way to reach the same results, don't you? ;)

Yes, I do realize the difference and I am only joking!

[Soulhunter]

Yes, you still have a small albeit considerable efficiency-gain of about 3% there but the merit of that for quality is perhaps questionable.
Yup, I absolutely agree with you in this point !!!

The quality gain is indeed questionable...




But the decisive thing is, if I would encode that clip from Matrix Reloaded to 20 MB in two-pass, the clips with no b-frames and that with your b-frame quantizer settings would feature an average p-frame quant of around 3, while the average p-frame quant of the clip with standard settings would still be 2.
So, it would be Q2 P-VOP + Q4 B-VOP -vs- Q3 P-VOP + Q3 B-VOP !!!

For me a Q3 P-VOP looks not more disturbing than Q4 B-VOP...

But of coz, its all a matter of taste/preferences !!!




And I could see that difference very well. And so could anyone, even with less costly equipment and a more casual look. So for any purpose where filesizerestriction of any kind exist (and this almost means 'always'), I would not recommend using ratio=1.00, offset=0.
True, me wouldnt recommend it as well... Imo its a waste of bitrate! If you have bits to spare, its better to use a CQM with lower frequency cut-off... But like I wrote before, ratio/offset=1 settings give imo better results than the defaults, at least when your aim is constant quality (visually). Of coz everybody should check/test this on his own!


Bye

[Sagittaire]

IMO quality with Bframe (qN for I,P and B) is very more important than without (qN for I,P) ... with VHQ or without VHQ.

[stephanV]

nope. i've tested many different sources but haven't found any discrepancies. it's just a 'rumor', as soulhunter pointed out. even b110 provides some, mostly negligible, gain in compression.
Actually I just posted a link to an old test of me with does indicate that, so no rumor at all.

when i'm testing i always go through a 000,110,111,121, ... sequence. and (believe or not) always (ok, in most instances) found 111 giving the highest quality metrics (w/significant gain in compression).
Those were comparisons at constant quantizer, it would seem. How about testing at identical filesize?

That comment made me frown a bit, its highly unlikely that at constant quant 1/1/1 would do better than 1/1/0 or no bvops with metrics. Therefor I extended my test a bit:

Xvid-1.1.0beta2, all settings default unless specified, Singe pass Q2 (first 1500 frames of Saggitaire's HP2 trailer)


| bitrate SSIM0 PSNR (avg)
--------------------------------------------------------------------------
no bvops, VHQ4: | 2755 kbps 89.64 46.7475 dB
bvops(1/1/0), VHQ4, no b-VHQ: | 2727 kbps 89.37 46.6376 dB
bvops(1/1/0), VHQ4, b-VHQ: | 2684 kbps 89.31 46.6042 dB
bvops(2/1/0), VHQ4, b-VHQ: | 2690 kbps 89.29 46.5947 dB
bvops(1/1/1), VHQ4, b-VHQ: | 2265 kbps 88.10 45.9830 dB


As you can see 1/1/1 is clearly worse than there rest with CQ. I think if you target a bitrate 1/1/1 would certainly be better though. (Not gonna try it out though). Of course bit rate is considerably lower too.

Anyway, this is all a bit over my head, I would never do anything near a Q2, as I don't really consider that compression. :)

[Teegedeck]

I guess we could label those like LAME presets for XviD 1-pass (if we assumed encoding at low quantizers):

1/1/0, identical to or worse than 0/0/0 = insane
1/1/1 = extreme
2/1.63/0 = standard
2/1.5/1 = medium*

Where standard is equal to extreme at low quants and equal to medium at high quants. Nice job, Didée!

:)

*: Note: everything 'standard' or better should be undistinguishable from the original, medium would be distinguishable in some cases.
Edit: I'd also like to mention that Koepi's old recommended settings (2/1.5/0.75) do something similar, at least for quant=2, as 2/1.63/0.

[Soulhunter]

@ Didée

Now thats tricky... ^^

Not thought of this way before !!!


@ Teegedeck

Presets would rock...

Same as pre-implemented CQMs !!!

But I fear we will never see this in XviD... :\


Bye

[Teegedeck]

Nope, naw, never. :(

We depend on len0x for any improvements on ease-of-handling.

[Harley Quin]

Sorry to interrupt your quality discussion, but I simply have to oppose to this "rumour"-thoughts...

With XviD 1.10 b2 koepi I encoded 40 min of "The Rock" at 688x288 and a quantizer of 2 at defaults.
Yes. Defaults.
Hey, we XviD newbies are regularly told to "stick to the defaults". Guess what, I did!
In this case it would especially mean: h.263, VHQ:1 and no VHQforB.
FileSize Avg p-frame Avg b-frame
Quantizer 2, no b-frames 478 MB 8.350 ---
Quantizer 2, B (2/1,0/0) =q2 482 MB 11.464 6.291
Quantizer 2, B (2/1,5/1) =q4 354 MB 11.464 2.341
Then I took the first 90 min of "The Usual Suspects", a very different film in terms of compressibility, with same frame-size and same default options:
FileSize Avg p-frame Avg b-frame
Quantizer 2, no b-frames 662 MB 5.052 ---
Quantizer 2, B (2/1,0/0) =q2 668 MB 7.093 3.984
Quantizer 2, B (2/1,5/1) =q4 468 MB 7.093 1.537
Undeniably a b-frame is smaller than a p-frame. That was never the point. The question is, whether the use of b-frame with the same quant as p is worth it. It's definitely not a rumour, "q2 b2" at defaults increases filesize compared to "q2 no b", at least for these two very different movies. The increse of p-size eats up the smaller b-size.

Ok, let's leave defaults.
With the 90 min of "The Usual Suspects" I changed the settings to VHQ:4, +VHQforB and only 1 consecutive B-frame and got these results:
FileSize Avg p-frame Avg b-frame
Quantizer 2, no b-frames 640 MB 4.877 ---
Quantizer 2, B (1/1,0/0) =q2 606 MB 6.049 3.140
Quantizer 2, B (1/1,5/1) =q4 481 MB 6.049 1.151
Here you get a gain with using b-frames without reduced quantizer. But... it's slight, and has to be well considered against the hassle that b-frames (and packed bitstream) can bring concerning standalones, VirtualDubMod and ffdshow.

Soulhunter, I said in my first post in this thread, that:
"the reduction in size, which comes with the use of b-frames, has its cause mostly due to the reduction of the quantizer, not the bi-directionality itself".
Bi-directionality alone lets you spare 34 MB, the reduced quantizer gives you another 125 MB.
Please tell me, what's "not true" about this?

Greetings

[stephanV]

It's very interesting to see in your results that the strain on p-frames (size increase) is independant of b-frame ratio/offset, which means the filesize gain completely has to come from the b-frames themselves. With that in mind a higher quant only starts to make more and more sense.

[Sagittaire]

yes but see the very interessing stephanV test: if you use bframe with 1.00/0.00 for ratio/offset you decrease quality. If you use real quant encoding (constant quant for I,P and B) bframe must be desactived.

Bframe are really interessing only if you use ratio/offset and encoding in constant quant with ratio/offset for bframe are not really a "constant quality" encoding in this case: average quant for low motion are very higher than average quant for high motion. If you want real constant average quant you must use 2 pass with modification for RC variability ...

[Soulhunter]

Bi-directionality alone lets you spare 34 MB, the reduced quantizer gives you another 125 MB.
And Im sure with even higher B-VOP quants you could get a 250MB reduction...

But unlike raising quants, the bidirectionality-gain causes no visible quality loss !!!



Please tell me, what's "not true" about this?
Well, It wasnt really "not true" but I knew other ppl would jump on !!!

So I saw my chance to prove that this "rumor" is not entirely true...

Sorry for abusing your post for this, really... ^^;


Bye

[Sharktooth]

B-frames are a sort of HVS optimization. They're meant to spare bits in places where the human eye cant spot details.
Q2 for b-frames nullifies that idea so i discard it by default :D.
However if you're gonna watch your movies frame by frame then good luck with q2 for b-frames...

[philippas]

Thanks for all your replies :)
I usually now encode my movies to be stored on dvd, 2 or 3 per dvd so quality is at greater importance for me.

For me, bframe quant=4 (1,1.5,1) introduce alot of blockiness.
bframe quant=3 is almost identical with quant2 for my eyes and the compressibility gain is good.

I just did a first pass with B-VOPS 3-1-1 and using VHQ1(bframes also)_QPEL_ CHROMAMotion_Turbo and here are the results:



MIN MAX AVG Total size(KB)
I 1,936 91,579 22,820 44,904
P 98 78,838 13,643 692,095
B 8 34,868 4,035 320,153




next the same encode with B-VOPS 3-1-0 and with no B-VOPS

[yaz]

Those were comparisons at constant quantizer, it would seem. How about testing at identical filesize?i never make constant quant encodes. another thing i haven't mentioned :sly: . my tests were all 2pass. as i do sg philippas outlined (2-4 films onto a dvd, so fitting sg to a certain size) constq is meaningless for me.

1/1/0, identical to or worse than 0/0/0 = insaneno. they are not identical (if 000 means off). just think of the 'n-vop discrepancy' discussed in an other thread. or just test 000(=off) against 0xx(on but no placement)

@stephanv
u're just reinventing spanish vax ;) it was told many times that using b-vop w/constq is suboptimal. (e.g., see the comments of sagittaire and sharktooth.)

@didée
the 'smooth operator' i like :D :D :D
however ... i distilled from my tests that allowing bq to get too far from pq is risky. it may have quite annoying effects crearly visible on steady scenes especially when the avgq is 'high'. that's why i prefer offsetting over multiplication.

@harley quin
yep ... 'stick to the defaults' ... if u follow this forum u may have noticed that almost all the default parameters were questioned. usually by people who know what they are talking about definitely. sad, but that's what we have. the onyl u can do is to go on to find your own way ... ehrm ... i mean, finding your favourite settings ;)

@all
all(!) of our testings have limited validity by nature. think of the great variability imposed by the parameters to be tuned in xvid. (there are at least 5 further parameters affecting the results of an encoding w/a fixed b-vop setting) multiply it w/the variability of the 'nature' and quality of the sources. then multiply it w/the variability introduced by cqms. and remember the different encoding modes (const q, fixed size, cbr, abr, whatever br) and ...
in such instances what a healthy human mind does is reduction. selecting some factors to be handled and letting the others at 'robust' defaults. and it's the point where personal preferences kick in.
i like crispyness. most people don't like it but i do. that's a definite preference of mine. i try to maintain this as much as i can, so, i don't set anything against this; no too high avg q, no too much b-vop, no flat matrices, aso aso. it concludes to b111 very easily ;)

my reading of b-vop is quite simply (as usual ;) ) : it provides significant gain in compressibility wout loosing quality. it may improve quality slightly(!) if i (u, we, ...) am (are) willing to sacrifice some of this gain. again, b111 fits this 'definition'.

i don't state b111 is 'the' setting but i do state it's a quite robust one and applicable for many different sorces safely.

the bests
y

f...ing typos :angry: sorry for that

[Didée]

Just a note for those that are not too deep into maths:

"0.62" is the "golden cut" value ( 1/x = 1+x ), which is e.g. proven to deliver geometric proportions perceived as "harmonious" or "beautiful". One more reason why I like the mentioned Bframe setting of x/1.63/0.

(1.62 vs. 1.63 makes a change at Pq = 8 --> Bq = 12 vs. 13)

[stephanV]

@yaz: I assume that with forcing b-frame quants to 2 (topic title) CQ is exactly what is discussed here... i cant believe anyone would do a 2-pass when he forces all quants to 2.

As for the statement that bidirectionality of b-frames has less significance to compression than the higher quants normally used with b-frames (which you called a "rumor") that has been proven sufficiently now, and can only be done by CQ. So what you call "spanish wax" now seemed not to so clear to you when you entered this thread. :)

[Mug Funky]

It's very interesting to see in your results that the strain on p-frames (size increase) is independant of b-frame ratio/offset, which means the filesize gain completely has to come from the b-frames themselves. With that in mind a higher quant only starts to make more and more sense.

this is logical though - "strain" on p-frames is simply a matter of motion-compensation becoming less effective. the strain on p-frames is dependant on the number of consecutive b-frames and the motion complexity of the source.

a p-frame can't motion-compensate off a b-frame, which means it's got to motion compensate off the previous p-frame (or i-frame) which can be several frames behind it. the image will have changed a fair bit in those frames, and the p-frame will therefore have to use more bits on the DCT part.

it could save a lot of bits to use lots of b-frames on a source that hardly moves at all, or moves predictably (like clean 2d-CG anime, or motion-graphics), but on an action movie with loads of film grain there'll be jack-all gain from using b-frames with the same quant as p-frames.

i think AVC changes this situation somewhat though - the rules are different.

[Soulhunter]

As for the statement that bidirectionality of b-frames has less significance to compression than the higher quants normally used with b-frames...
But a compression gain alone means nothing!

You have to take the resulting quality into account...

Otherwise we would all use offset values of 20 or so, no?


Bye

[MeteorRain]

well, i want to say somethings
i have already been used to encoding files using bf(4/1/0), or sometimes even small quantilizer.
imho, the b frame is used to replace p frame when it's not so motive, and thus saves hugh amounts of bitrate(filesize of coz)
when in the same quantilizer, if, pf, bf gives the same quality, so the more you use bf, the more bitrate you will save. and there4, the saved bitrate can be add up to if and pf to gain a better quality, or just don't do anything and save the harddisk space.
imo, a quantilizer at a constant (or say not very dynamic), eg 2,3,3,2,3,3,3,2,2,... is better than 2,2,5,2,5,2,5,2,5... and ofcoz have a better output quality.
well, my own opinion, disscusion welcomed!

[Soulhunter]

@ MeteorRain

Yup, thats what i meant in one of my earlier posts...

Ratio/Offset=1 gives imo more constant quality than the defaults !!!


Bye

[stephanV]

But a compression gain alone means nothing!

You have to take the resulting quality into account...


I never denied that. With metrics I have already already proven that even b-frames at the same quant might be worse in quality than p-frames...

So a. you have little or no compression gain
and b. possibly less quality

when in the same quantilizer, if, pf, bf gives the same quality, so the more you use bf, the more bitrate you will save.
Well, they might possibly not give the same quality and dont always give a compression gain. a setting of (x/1/0) just makes no sense at all. Or better said, in a CQ scenario b-frames make no sense at all.

[Soulhunter]

I never denied that. With metrics I have already proven that even b-frames at the same quant might be worse in quality than p-frames...
Yup, same results here with most stuff I tested !!!

Slight compression gain, but also a slightly worse metric score...

Though for ratio/offset=1 -vs- defaults the things look a bit different !!!


Bye