Which is the optimal motion search range in x264 and x265?

[sophisticles]

Of course it's back assward. Consider a video that gets 0.999995 and another video that gets 0.999999. By their calculations:

((1-0.999995)/(1-0.999999) - 1)*100 
((0.000005/0.000001)-1)*100
(5-1)*100
4*100
400

The bad video is 400 percent worse than the good one! It must look like total crap, huh? No, they're both nearly perfect.

If your boss told you he was giving you a 400 percent pay raise would you expect the calculation to be made like this?

PSNR and SSIM are measured in decibels and thus use a logarithmic scale, not a linear one.

[Selur]

for further investigation into the formula, it was posted by Dark Shikari a while back, see: https://forum.doom9.org/showthread.php?t=150118

[El Heggunte]

<Off-Topic>

-- Doom9's SSL/TLS implementation is b0rked, so please avoid posting https links to his forum;

-- (to whom this may interest) Dark Shikari has renamed himself to «Fiona Glaser»

</Off-Topic>

[Selur]

Doom9's SSL/TLS implementation is b0rked, so please avoid posting https links to his forum;

will try to remember

[hello_hello]

Of course it's back assward. Consider a video that gets 0.999995 and another video that gets 0.999999. By their calculations:

((1-0.999995)/(1-0.999999) - 1)*100 
((0.000005/0.000001)-1)*100
(5-1)*100
4*100
400

The bad video is 400 percent worse than the good one! It must look like total crap, huh? No, they're both nearly perfect.

If your boss told you he was giving you a 400 percent pay raise would you expect the calculation to be made like this?

PSNR and SSIM are measured in decibels and thus use a logarithmic scale, not a linear one.

What is a decibel in respect to video quality? I have no idea.

I can get my head around converting amplitude ratios to a percentage ie

-12dB = 25%
- 6dB = 50%
+6dB = 200%
+12dB = 400%

without my brain hurting too much, but the SSIM formula is beyond my comprehension.

[sophisticles]

http://en.wikipedia.org/wiki/Decibel

http://en.wikipedia.org/wiki/Structural_similarity

As you will note SSIM value of an image ranges between -1 and 1, with a value of 1 representing 2 identical sets of data and a value of -1 representing 2 completely dissimilar sets of data. SSIM is only applied on luma, i.e. the brightness of an image and thus for all practical purposes SSIM measures the difference in brightness of 2 images.

Once you understand that you understand why it's measured in decibels and follows a logarithmic scale, in much the same way decibels are used to measure sound pressure along with acoustic intensity they are also used to measure luminous intensity.

For our purposes that means that when we measure the SSIM of an image relative to a reference image, a value of 1 would mean that they have the same brightness or luminous intensity and value of -1 would mean that each 8x8 pixel window of each image has a different brightness. We need to keep in mind that this metric does not tells us whether comparison image is brighter or darker than the reference image, only that they do not have the same luminous intensity and we can't really use it to determine which looks better.

It an interesting metric for comparing 2 images.

A more interesting note is that both x264 and x265 which allow one to set a "tune ssim" option, this results in

--aq-mode 2 --no-psy

which means if you want the closest luminous intensity between your source and encode you should disable psy and set AQ to auto, which would lead one to conclude that both x264 and x265 implementation of psycho-visual optimizations muck with luminous intensity, in the attempts to improve perceived quality.

Whether this is a good thing or a bad thing is up to each individual to determine.

[hello_hello]

which means if you want the closest luminous intensity between your source and encode you should disable psy and set AQ to auto, which would lead one to conclude that both x264 and x265 implementation of psycho-visual optimizations muck with luminous intensity, in the attempts to improve perceived quality.

Whether this is a good thing or a bad thing is up to each individual to determine.

I've never paid any attention to SSIM or PSNR measurements myself (which is why I know nothing about them), because I've never had much faith they translate into quality measurements as humans would perceive quality. I recall.... and amazingly found...... an old thread where someone was wondering why SSIM showed transcoding produced better quality than re-encoding. http://forum.doom9.org/showthread.php?t=112034

Anyway, thanks for the info. I'm still not sure I understand the dB relationship to video quality though. The SSIM formula mentions "dynamic range", but..... I still can't translate any of it into something my brain can understand as a difference between two results. The equation posted earlier which shows a difference of 400%. I can't get my brain to make sense of that. I'm not saying any of it's wrong.... just that I don't understand it yet.

[Mephesto]

Of course it's back assward. Consider a video that gets 0.999995 and another video that gets 0.999999. By their calculations:

((1-0.999995)/(1-0.999999) - 1)*100 
((0.000005/0.000001)-1)*100
(5-1)*100
4*100
400

The bad video is 400 percent worse than the good one! It must look like total crap, huh? No, they're both nearly perfect.

If your boss told you he was giving you a 400 percent pay raise would you expect the calculation to be made like this?

Wow, jagabo. For someone that's the second most useful poster on this site you're really talking out of your ass on this point.

The guy's video that we were comparing was not a perfect vs. perfect encode but a crappy 0.98400 vs. a much better 0.99000. No video without hideous flaws gets an average score of 0.98400. A 0.99000 signifies really good quality. Anyone experienced with the metric knows this. You are pulling a straw man argument out of your ass.

Btw, when encoding NES speedruns, I've often gotten ridiculously high SSIMs many times like 0.99950 because all the static scrolling objects are perfect quality while only the sprites who don't move in just one predictable direction are visibly affected. Sometimes, the moving character is significantly smeared with artifacts while 99% of everything else is literally untouched quality. When increasing the --merange to 64, the quality of the moving character was significantly improved and the average SSIM moved up to 0.999980 or thereabouts. Yup, 300% quality increase and I agree with it because I definitely did notice being able to actually make out all the details of the main, moving object that most will be focusing on the entire video.
So what's your point exactly?

I've never paid any attention to SSIM or PSNR measurements myself (which is why I know nothing about them), because I've never had much faith they translate into quality measurements as humans would perceive quality. I recall.... and amazingly found...... an old thread where someone was wondering why SSIM showed transcoding produced better quality than re-encoding. http://forum.doom9.org/showthread.php?t=112034

Anyway, thanks for the info. I'm still not sure I understand the dB relationship to video quality though. The SSIM formula mentions "dynamic range", but..... I still can't translate any of it into something my brain can understand as a difference between two results. The equation posted earlier which shows a difference of 400%. I can't get my brain to make sense of that. I'm not saying any of it's wrong.... just that I don't understand it yet.

Play around with it and you'll get used to it.
General rule of thumb with SSIM:
0.99500+ = HQ
0.99000 = good
0.98500 = fair
0.98000 and below = bad

[jagabo]

Of course it's back assward. Consider a video that gets 0.999995 and another video that gets 0.999999. By their calculations:

((1-0.999995)/(1-0.999999) - 1)*100 
((0.000005/0.000001)-1)*100
(5-1)*100
4*100
400

The bad video is 400 percent worse than the good one! It must look like total crap, huh? No, they're both nearly perfect.

If your boss told you he was giving you a 400 percent pay raise would you expect the calculation to be made like this?

Wow, jagabo. For someone that's the second most useful poster on this site you're really talking out of your ass on this point.

My point was that the percentage difference stated this way was not an indication that the worse video had to look like crap. Obviously, you must consider the absolute values.

[Mephesto]

Of course it's back assward. Consider a video that gets 0.999995 and another video that gets 0.999999. By their calculations:

((1-0.999995)/(1-0.999999) - 1)*100 
((0.000005/0.000001)-1)*100
(5-1)*100
4*100
400

The bad video is 400 percent worse than the good one! It must look like total crap, huh? No, they're both nearly perfect.

If your boss told you he was giving you a 400 percent pay raise would you expect the calculation to be made like this?

Wow, jagabo. For someone that's the second most useful poster on this site you're really talking out of your ass on this point.

My point was that the percentage difference stated this way was not an indication that the worse video had to look like crap. Obviously, you must consider the absolute values.

0.98400 is not particularly good quality, so the absolute value is being considered.

[mzso]

<Off-Topic>

-- Doom9's SSL/TLS implementation is b0rked, so please avoid posting https links to his forum;

-- (to whom this may interest) Dark Shikari has renamed himself to «Fiona Glaser»

</Off-Topic>

So sex "change" or wtf? A tad weird.

[Selur]

If you want to know if Fiona Glaser had a sex change or why the name was changed, best ask the user directly. Picking up a 3+ month old off-topic comment seems more like trolling than anything else.
Fionas last commit to x264 (according to http://git.videolan.org/?p=x264.git;a=shortlog) was 2014-07-20 so no noticeable public activity.
-> If that confuses you, never start coding and reading the code of other folks. *gig*

[sophisticles]

If you want to know if Fiona Glaser had a sex change or why the name was changed, best ask the user directly. Picking up a 3+ month old off-topic comment seems more like trolling than anything else.
Fionas last commit to x264 (according to http://git.videolan.org/?p=x264.git;a=shortlog) was 2014-07-20 so no noticeable public activity.
-> If that confuses you, never start coding and reading the code of other folks. *gig*

He may or may not be "trolling" but it's certainly a question I asked myself. Jason not only changed his name to "Fiona" but he also changed the copyright on all his commits to all open source projects going back to when he/she first took over the x264 project.

But the "strange" doesn't end there, there was an old article where the author said something to the effect of "x264 lead developer Jason said...", when I was searching for something and happen to run across the same article it has been retroactively been changed to "x264 lead developer Fiona said...".

Perhaps the "funniest" thing about this whole thing is that not only did Jason close down Doom10 but in another forum where "Dark Shakari" has an account the avatar used was a left half picture of Jason's face with a girl's hairstyle and makeup and believe me I know what Jason looks like and that was him/she.

Jason has been out of the closet for almost a decade now, it certainly seems like he/she is in "Pre-Op" mode, maybe he/she finally made enough money from x264 to be able to afford a sex change.

In the end it makes no difference to me, whatever makes him/her happy but it's very strange that he/she would do this in such a public way.

[Selur]

he/she would do this in such a public way.

Was there ever another alternative? I mean given that she is/was the main face behind x264 for quite some time.
-> Only wish her the best.

[sophisticles]

he/she would do this in such a public way.

Was there ever another alternative? I mean given that she is/was the main face behind x264 for quite some time.
-> Only wish her the best.

I suppose it had to be public but Jason/Fiona should if he/she wanted to be public about it then just make an announcement and be done with it; this way is like public but in a weird way, it caught everyone by surprise, even the members over at the Doom9 forums, like Lord Mulder, who have known Jason for years were caught off guard when he/she retroactively changed the copyright on all his/her commits with the -force option.

While I wish him/her the best, I know he/she has a tough road ahead, in the U.S. one has to live for 2 years as a member of the opposite sex before they can have the gender reassignment surgery, they have to go for a psych evaluation, then there's all the legal matters with changing your name and sex on driver's license, social security, the criminal background checks to make sure you're not trying to escape prosecution, it gets very expensive between legal fees and medical bills.

None of that is any of our business of course but what does concern us is that it seems like Jason/Fiona has not been the main developer of x264 for a while, reading through the commits it seemed like a couple of Russian programmers had taken over the development and maintenance of the project. One has to wonder if this is a signal from Jason/Fiona and Loren that x264 is nearing it's EOL. As poisondeathray pointed out in another thread, that new aq mode that was recently committed to main has been around for 5 years and was considered experimental all this time, now all of a sudden it gets pushed to main.

I also wonder if this means that Jason/Fiona has decided not to work on daala, it was rumored for a while that both he/she and Loren were going to be working on that codec, at the very least that they had expressed a desire to do so, it seems like it's all up in the air now.

[Selur]

Don't think that she dropped x264 and switched to daala (no commits there either), but we will see.

it seemed like a couple of Russian programmers had taken over the development and maintenance of the project.

Most of them have been committing for years to x264, I guess it's just that the old core team doesn't seem to commit much the last half year.

-> think we should really switch back to x264/x265 and me range

[mzso]

...

Ugh.

Don't think that she dropped x264 and switched to daala (no commits there either), but we will see.

it seemed like a couple of Russian programmers had taken over the development and maintenance of the project.

Most of them have been committing for years to x264, I guess it's just that the old core team doesn't seem to commit much the last half year.

-> think we should really switch back to x264/x265 and me range

Maybe it's not worth the effort anymore. I can imagine that after this many years of optimization the capabilites of the format are near completely exploited, and there isn't much room for improvement in optimizing speed. So the efforts would have greatly diminished returns.

[_Al_]

yes, life goes on, he has helped to lots of people already, I'd encourage him/her (if it is not a joke) to go on, tackle something else, or to just enjoy a glass of wine every evening ...

[sophisticles]

yes, life goes on, he has helped to lots of people already, I'd encourage him/her (if it is not a joke) to go on, tackle something else, or to just enjoy a glass of wine every evening ...

Definitely not a joke, the copyrights on the commits where changed from the Dark Shakari account, retroactively, with the -force command, Doom10 was closed down and I did she that avatar and there have been rumblings for a while within certain circles.

Regardless, I tend to agree that we have probably seen AVC encoder in general taken to the limits of quality and speed, there's probably little they can do to improve x264 now because the code base is so "baked in" that trying to correct it's short comings would require substantial rewrites.

I would rather see anyone interested in working on x264 to fork the x265 code base and have a variant not in the control of a corporation or work on a gpu accelerated version of f265.

[mzso]

I would rather see anyone interested in working on x264 to fork the x265 code base and have a variant not in the control of a corporation or work on a gpu accelerated version of f265.

Why does it bother you? Corporations are more equipped to manage projects and keep things going (by paying employees)
Splitting up the devs would most likely only have negative consequences, in my opinion.

[sophisticles]

Why does it bother you? Corporations are more equipped to manage projects and keep things going (by paying employees) Splitting up the devs would most likely only have negative consequences, in my opinion.

I would beg to differ, there are many open source projects that are managed just fine without any corporate input whatsoever, some examples:

KDE
GNOME
The Linux Kernel
BSD
xvid
vlc
x264, which despite the x264LLC offshoot is not corporate controlled

The list goes on. When a project is in the control of a big corporation it's development is dictated by bean counters, they do what is best financially not necessarily best for the development of the project.

[movmasty]

<Off-Topic>

-- Doom9's SSL/TLS implementation is b0rked, so please avoid posting https links to his forum;

-- (to whom this may interest) Dark Shikari has renamed himself to «Fiona Glaser»

</Off-Topic>

Uhmm, ok, but, who is Rebecca Black ?

BTW, Shikari was already a girl name

-> Only wish her the best.

quote

[El Heggunte]

0x00007f00