B-frames

[Turbonutter]

Hey there.

I was doing some encoding tests on The Final Flight of the Osiris (The Animatrix chapter 1), using the following settings:

6 - Ultra High
H.263
VHQ 2
300
1
qpel and chroma motion
2
150
100
100

The interesting thing was, I found the quality to be better (more detail, less macroblocks) when I had b-frames disabled than when I didn't - bitrate 864kbps (but using 2-pass int). Is there something I'm not getting here? I thought B-frames were supposed to improve quality.

Anyway, I'm going to try a section of High Fidelity later to see if I get the same results, but can I ask other people's experiences with b-frames?

[OUTPinged_]

Try with 1 b-frame.

[Gaia]

Qpel for animation is useless.

Change your b-frame settings to 150,100,0. B-frames improve quality a lot but i don't have any exeperience encoding animation. B-frames and animation migh be totally different thing. I don't have any experience encoding animation.

[bond]

what xvid build did you use?

try 2b-frames with threshold 0

[Koepi]

I set offset to 75 in the meantime btw., it seems to give even better results and restores the possibility to compress huge 1st passes higher.

For example, with 2/150/100/0 you can compress from 900mb 1st pass to 600mb 2nd pass, with 2/150/75/0 you can do 1100mb to 600mb with acceptable quality. (Ok, depends on the source, but just an observation i made.)

Regards
Koepi

[entropy]

Originally posted by Turbonutter
Hey there.

I was doing some encoding tests on The Final Flight of the Osiris (The Animatrix chapter 1), using the following settings:

6 - Ultra High
H.263
VHQ 2
300
1
qpel and chroma motion
2
150
100
100

The interesting thing was, I found the quality to be better (more detail, less macroblocks) when I had b-frames disabled than when I didn't - bitrate 864kbps (but using 2-pass int). Is there something I'm not getting here? I thought B-frames were supposed to improve quality.

Anyway, I'm going to try a section of High Fidelity later to see if I get the same results, but can I ask other people's experiences with b-frames?

I found the same result. I think it looks better without b-frames, but the size does go up.

[Prettz]

Originally posted by Gaia
Qpel for animation is useless.

I could beg to differ. Some anime looks so similar to live action (with the ultra-detailed backgrounds, CG special effects, and lighting effects and such applied to the cel animations) that the rules for normal movies apply rather than those for animation.

[Turbonutter]

Heh yes, anime like Ghost in the Shell is more than animation - which is what I encode.

OT; Anyway, Final Flight is actually a Final Fantasy style CGI movie, very detailed and also has lots of motion in places. I find it quite easy to use as a comparison as the original is a perfect reproduction, it doens't contain artefacts and annoyances while keeping many of the qualities of actual film. Incidentally The Spirits Within would be much better for this, but I don't have that DVD.

Anyway, back to B-frames. I understand that on a pure quality basis B-frames probably are very good, but I'm more interested in qualty for a given filesize, DVD-backup style. I think under these circumstances, quality degrades with B-frames, right?

Incidentally, comparing two rips of the 10 minute Final Flight compressed to 58mb, both with B-frames, setting threshold to 100 is undoubtly better quality than 0.

OT: As for qpel with animation, well, I have to encode some Family Guy episodes tomorrow, so I'll do a comparison of qpel vs. no qpel on that. I'll also encode a segment of Ghost in the Shell under the same conditions and post the results in a new thread, in the interests of knowledge.

EDIT: XviD build is Koepi 14052003-1

[Gaia]

Anyway, back to B-frames. I understand that on a pure quality basis B-frames probably are very good, but I'm more interested in qualty for a given filesize, DVD-backup style. I think under these circumstances, quality degrades with B-frames, right?

I don't understand what you mean but in my own encodes(full movie dvd sources) b-frames don't decrease quality. They improve it a lot...

[Turbonutter]

Well.. in my experience, quality degrades, but we'll see tomorrow. I'm going to do a straight comparison of -1/ and 2/150/100/100 on High Fidelity (614mb), and see what happens - I'll happily eat my words if the quality is better!

Maybe it depends on the movie?

[Nibor]

Why do you use such a outdated build like this :p ;)

IMHO Koepi's 24062003-1 build with B-frames settings 2/150/75/0 gives great results for me! In a given filesize (and 2 pass of course) b-Frames increases quality, but maybe it really depends on the movie... Or it's just a matter of taste!

Never did an anime encode, will start with Princess Mononoke :)

Ok, just my 2 cents...

[symonjfox]

Hey guys, maybe you don't know exactly what Animatrix is.
The final flight of osiris is NOT an anime, it's a 3D graphic cartoon. It's very detailed and there's a lot of motion, effects and detail in it. IMHO Qpel can help in this episode (well in other episodies maybe not ... who knows).

Back to topic:
I had a long discussion in other threads about B frames in High Quality encodes. Well me too I though that a very high bitrate movie without Bframes should look better. Instead, B frames are very useful. Maybe you should try 2/100/100/0 setting. Doing so, the B frame will have the same quantizer of a P frame that would be insered there, but B frame can refer to previous or next frame, so quality can be higher at same filespace.
Bframes in low bitrate encodings is a MUST!

[sysKin]

Originally posted by Turbonutter
Well.. in my experience, quality degrades, but we'll see tomorrow. I'm going to do a straight comparison of -1/ and 2/150/100/100 on High Fidelity (614mb), and see what happens - I'll happily eat my words if the quality is better! OF COURSE it will be worse with crazy settings like threshold 100. Why won't you try default first? 100 is a huge value (and I'm talking about the old build), and yes, it will look worse.

[Turbonutter]

Phft thresholds of 100 are recommended, and a threshold of 100 definitely looks better than threshold 0.

Anyway, the recent encode without B-frames also looks better, but I will disregard this in view of the threshold argument.

I use this build because it comes with gknot - iirc from visiting Koepi's website it's the latest development build, no?

Anyway, I'll dig out The Fifth Element and do some 10-minute 720 encodes after Family Guy (probably Friday) and settle this...

I'll be using the above settings, including qpel, if anyone wants to suggest some B-frames settings to use (and if someone can suggest a decent way of comparing qualities (http://forum.doom9.org/showthread.php?s=&threadid=56329), I'll let you know how it goes.

[Nibor]

Originally posted by Turbonutter
I use this build because it comes with gknot - iirc from visiting Koepi's website it's the latest development build, no?
No, it isn't! Go check Koepi's Site (http://roeder.goe.net/~koepi/xvid.shtml)..
IMHO Recommended B-frame settings in the newest build are 2/150/75/0
Beam me to the thread (http://forum.doom9.org/showthread.php?s=&threadid=56190) :D

[JasonFly]

A bframe encode will always lokks worse than a without one.
But using the defaut setting will give good results and the difference will not be much noticeable in playing(In most cases).

But that's true that i noticed some blocks in very fast motion scene in Final FLight of Osiris.(The scene was a fast one and a blue/dark one).

Maybe my bframe treshold was a bit too high(20 with the new build if i remember well)

@ Syskin:
Does the bframe decision take into account motion(I presume it does but what does it looks like)?

[sysKin]

Originally posted by JasonFly
Does the bframe decision take into account motion(I presume it does but what does it looks like)? No, not motion. It looks like motion, but it's a side-effect, which I'll try to counter soon because panning = lots of motion AND good for bframes.

The bframe decision takes into account the total texture data that needs to be coded after motion compensation. The more new texture -> the sooner we need a pframe, the less new texture -> the more bframes.

Radek

[Turbonutter]

So what do B-frames actually do? I know they can take data from frames before AND after them, but how does this translate to advantages? Better texture quality?

[gldblade]

B-frames are more heavily quantized, which means they are smaller than normal P-frames. By taking up less space, more bits are made available to P-frames and I-frames, translating to higher overall quality.

[Turbonutter]

Okay, well in the absence of my Animatrix disc, I did some testing on Chapter 7 of The Fifth Element, which I think has to be one of the most varied 7 minutes in any film - you've got nice close ups (lots of lovely skin textures), some really fast motion sequences, and plenty of those all-important smoke effects that codecs seem to hate.

The AVS is 7675 frames @ 25fps, 720x294, encoded using 2-pass int, 31423kb, MSP 6 VHQ 4 H.263 qpel, chroma motion and optimisation. Encoding was circa 10fps on 2075mhz AMD, 333mhz DDR.

Alright, using the new codec and 2/150/75/0, the quality is much, much better, especially on the motion sequences.

http://www.sine-wave.net/stuff/xvid/tfe-5990-non.png
http://www.sine-wave.net/stuff/xvid/tfe-5990-2.150.75.0.png

I know these aren't exactly the same frame but I assure you the quality is representational.

The quality may be slightly higher on the close up shots without B-frames, but I really can't tell.

-1 first size was 139mb
2/150/75/0 first size was 115mb
2/150/100/100 first size was 90mb. The quality of this was crap.

If anyone wants any more testing/screenshots/data done on this sequence I'll keep all the files until tomorrow (Sunday) evening. I'm on 56k so I won't upload anything bigger than about 5mb, so I guess I can up a sequence if you wish.

Incidentally SBC's quality in this particular frame was much, much better, however it did overshoot the filesize by 8.5mb giving it a big advantage, and the general quality throughout the film is worse, with a constant macroblock artefact across all the scenes. I'm a bit worried at how SBC has suddenly decided to overshoot everything. It was four times faster though.

[bugsan]

why bframes give a lower PSNR ?

a simple subtract(original, encoded)
xvid default+bframe1
ffdshow 030103

http://perso.wanadoo.fr/bugsan/2240.png
http://perso.wanadoo.fr/bugsan/2241.png
http://perso.wanadoo.fr/bugsan/2242.png
http://perso.wanadoo.fr/bugsan/2243.png
http://perso.wanadoo.fr/bugsan/2244.png

[Turbonutter]

Interesting you should mention PSNRs. I found the encodes without B-frames gave higher PSNRs overall, but I must say the B-frames one did appear visually better.

[gldblade]

It's important to note that PSNR is not an accurate measure of visual quality. Just because PSNR goes down does not mean that quality goes down.

B-frames have lower PSNR because they are more highly quantized than other frames.

[Assault]

@ gldblade, Turbonutter

What's even more important: If you use 1-pass quantizer mode your b-frames encode will have a lower PSNR value than your encode without b-frames but what's more important it will also have a lower filesize. That's the reason why in a 2-pass encode PSNR is higher for encodes WITH b-frames than for encodes WITHOUT. (the filesize gain is bigger than the quality loss because of b-frames) ;)

Assault