Psy RDO: Official testing thread (version 0.6 out!)

[Shinigami-Sama]

Originally Posted by tetsuo55

Maybe the combination of patches and settings fachman uses is leading to unexpected results (higher quality encodes without increasing bitrate)

I agree with DS that we need samples, single screenshots cannot tell us anything about the temporal quality.

If their really is a huge improvement when compared to regular settings maybe DS could find the cause and create a patch so those insane settings are no longer needed.

thats like saying randomly changing mpc-hc's code will make my old ISA trident 1mb videocard do dxva on 4k lossess videos...

[fachman]

Originally Posted by IgorC

Excuse me but who is Komisar? That guy who test on a very fewest videos looking for SSIM and OPSNR values and not comparing them visually that much. And another guy like nico who "helps" him to tune x264 his way and thinks that psyrdo and psytrellis are totally waste of time.

I would stay away from these underground redone algos something like VAQ2modified, VAQ3 ..generally VAQN+1.

Dear IgorC

Although I have no idea how Komisar did his tests, for the future I would consider think twice before you will say something, because testing on "very fewest videos" is normal thing which you would know if you would known how Dark Shikari is doing his tests http://x264dev.multimedia.cx/?paged=2. Unfortunatelly we are facing reality here.

Also I could bet you do not live with Komisar or Nico to know how they did their tests. Do you????
Althought SSIM and PSNR are not perfect I would rather trust them than someone elses "opinion". Why??? Because human opinion can change. On the morning you can like the picture. On the afternoon you will not, and on the evening you could barelly look at it, but Quality metrics will stay the same.
While I do see a big sence in PSYRDO it is true that it lowers SSIM and PSNR. Thus I would suggest another metrics which compares complexity of the scene.
IMHO the best picture would give the picture which balances those 3 metrics.

[DarkZell666]

Originally Posted by fachman

Althought SSIM and PSNR are not perfect I would rather trust them than someone elses "opinion". Why??? Because human opinion can change. On the morning you can like the picture. On the afternoon you will not, and on the evening you could barelly look at it, but Quality metrics will stay the same.
While I do see a big sence in PSYRDO it is true that it lowers SSIM and PSNR. Thus I would suggest another metrics which compares complexity of the scene.
IMHO the best picture would give the picture which balances those 3 metrics.

Are you so lunatic ? I'm not Also, you seem not to even trust your own opinion, nevermind someone else's ... are your eyes so damaged ? (no offense, I'm just intrigued by the fact you rely more on numbers that on what you see )
Moreover, you seem to omit the fact that (and this has been repeated too often already for you not to have noticed imho) higher metrics DON'T ALWAYS mean higher perceptual quality.
Proof ? : All the recent psy optimisations kill PSNR and/or SSIM but 90% of the people hanging around here have noticed visual improvement. So if you mean to say you're in the 10% that don't, well, just say so, we won't blame you for it for the same reason we don't blame PSNR and SSIM for being tricked by psy improvements, it's in their nature

[CruNcher]

@Darkzell666
Even PsyRD is a two handed sword because you seeing a improvement in a very high compression area wich is RD also you can't say that PsyRD allways have this effect for every source and every situation, also seing a effect on 1 or 2 frames doesn't mean consistent results and then some people prefer to take Speed/Their Subjective Quality (for PsyRD that is it biggest bonus it comes for free with RD, so it's also profile indepedent tough your investment is a 20% slowdown) against it.
Also some are even taking into account @ wich bitrate it becomes useless or even backfiring @ you , all of this is much more complex, then what people might only see in 1 bitrate and compression scenario here it's true for their situation the're in but it didn't have to be true (or as good) for others,if you dont care about speed or balance @ all but optimal quality every of this ofcourse becomes invalid for you, but if you try todo Psy in some more balanced way you have todo alot more compromises. I'm also skeptical about fachman his experience tough i would never say it isn't true for the situation he's in, tough im very skeptical and even would agree with Dark here changeing the RC in that way seems to call for problems in a consistent visual overall result (runtime), but he is useing Komisars build so --aq-mode 3 if i see right and i dont have to much experience visualy yet with it as im more working around with --aq-mode 2 non scaled like in Komisars build currently. Btw all of this becomes even more complex if you add more restrictions like VBV to the pool

[fachman]

Originally Posted by DarkZell666

Are you so lunatic ? I'm not Also, you seem not to even trust your own opinion, nevermind someone else's ... are your eyes so damaged ? (no offense, I'm just intrigued by the fact you rely more on numbers that on what you see )

Well, you know what I am intrigued where did you get the idea I do not trust my opinion??? Because if you are really not offensive, just intrigued you should find it easily

Moreover, you seem to omit the fact that (and this has been repeated too often already for you not to have noticed imho) higher metrics DON'T mean higher perceptual quality.

Are you sure there wasn`t the word ALWAYS somewhere between the words???

Proof ? : All the recent psy optimisations kill PSNR and/or SSIM but 90% of the people hanging around here have noticed visual improvement. So if you mean to say you're in the 10% that don't, well, just say so, we won't blame you for it for the same reason we don't blame PSNR and SSIM for being tricked by psy improvements, it's in their nature

Oh my god, where did you get it that I do not noticed PSYRDO visual improvement.
The reason I trust more quality metrics is because someone elses opinion are very often based not what IT IS, but what it SEEMS to be.

[Ranguvar]

Ah, but in a field where all that matters is human perception, what seems to be is all that matters

Metrics can report all day that a video is low or high quality, but none take into account the inaccuracies of the human eye, and the best encoders are the ones that exploit these inaccuracies. Compress where the eye doesn't notice, don't where it doesn't. Etc.

A good old double blind test with a large survey group is by far the best way, IMO, to measure a feature's worth. Metrics can be useful short-term for optimizing a few things quickly, like b-frame decision, but overall they are meaningless. Unless you can come up with, again, a metric that is like the human eye.

[kemuri-_9]

Originally Posted by kemuri-_9

hmmm... looking at the code from the patch more closely now, there's no longer a way to use normal trellis with psy-rdo in the psy-trellis inclusion patch,
is there anyway that an option could be added in a future release to be able to enable/disable psy-trellis so that psy-rdo could be used with standard trellis again (looking at you trellis 2)?
I think it would be helpful with testing/comparing...

So since you were already busy DS, i tried tackling this myself and the result is:
x264_psy_rdo.0.5+psy_trellis_01_r929_mod.diff

added a new --psy-trellis option to enable and control code usage for psy trellis
allowing it to be on/off similar to psy rdo (within its dependencies of trellis > 0 and psy-rd > 0)

of course this needs your review when you have the time to make sure i didn't fubar anything.

[DarkZell666]

The reason I trust more quality metrics is because someone elses opinion are very often based not what IT IS, but what it SEEMS to be.

The whole point of psy-rdo is that what ever IT IS, it SEEMS to be better than what it actually IS => Read : whatever the metrics (good or bad), it looks better.

And what's the matter with "someone else's" opinion you keep talking about : what about YOUR opinion ? What about YOUR eyes ? Forget the others, forget the numbers ... what do YOU see ? Which output do your eyes prefer ? You're putting the responsibility of what you percieve on "someone else", or "something else" (PSNR/SSIM). You seem to be saying "PSNR says it's better, so better it is". Am I wrong somewhere ? I'd be glad you correct me straight away !

Looking at an encode through metrics (which is what you apparently do) is like voluntarily watching a movie without the psy-enhancements. Do you get the idea ? It's voluntarily limiting your perception of the actual picture being displayed.

And yes, there was the ALWAYS word missing (post edited), I thought what I was getting at was obvious but I'll be careful next time .

[Dark Shikari]

PSNR 37.165 and 38.152 respectively, 300kbps.

Now stop talking about useless metrics.

[Razorholt]

Are you sure you are comparing the same type of frames?

[Dark Shikari]

Originally Posted by Razorholt

Are you sure you are comparing the same type of frames?

Yes, both are P-frames.

[Razorholt]

So, does it mean that we "may" compare frames as long as they are the same type? ... I mean, in order to compare encoded files.

[Dark Shikari]

Originally Posted by Razorholt

So, does it mean that we "may" compare frames as long as they are the same type? ... I mean, in order to compare encoded files.

Comparing frames is usually fine, but you have to watch out for cases where it might be misleading (for example, if ratecontrol gave significantly different distributions of bits between two files).

By the way, the first image is latest x264 with maxed settings, b-adapt 2, and psyrd+psytrellis, and the second is x264 r614 with maxed settings.

[MasterNobody]

Dark Shikari
And can you tell the SSIM for this samples?

[smok3]

a little transcode test, besides some funny marks on some faces (posterization enhanced?) looks really good imho;

http://somestuff.org/flashAVC/flvpla...-x1280y544.mp4

p.s. useless encoding time.

[Atak_Snajpera]

Yes I agree! Excellent quality @ 1.8 MBps.

[Dark Shikari]

Originally Posted by MasterNobody

Dark Shikari
And can you tell the SSIM for this samples?

I don't have the logs anymore; the SSIM was just about the same for both (negligable difference), around 0.95.

Originally Posted by smok3

a little transcode test, besides some funny marks on some faces (posterization enhanced?) looks really good imho;

http://somestuff.org/flashAVC/flvpla...-x1280y544.mp4

p.s. useless encoding time.

[Sagittaire]

Originally Posted by Dark Shikari

PSNR 37.165 and 38.152 respectively, 300kbps.

Now stop talking about useless metrics.

Well it's paradoxal here because codec is simply mathematical algo. If you say that the first algo is better than the second algo for eyes you can by definition produce other mathematical algo (metric) for show that ... isn't it?

[jethro]

psy-rd is that 'mathematical algo (metric)', I think

[lexor]

Originally Posted by smok3

a little transcode test, besides some funny marks on some faces (posterization enhanced?) looks really good imho;

http://somestuff.org/flashAVC/flvpla...-x1280y544.mp4

p.s. useless encoding time.

That is awesome (the encode I mean, the movie looks like it'll blow chunks). But did you use tv-pc colour conversion? The blacks don't look quite right, and for web playback the colours should always be converted to pc scale (since we can't have flash player post process in browser).