Optimum Resolution

[Jeff Mott]

I decided I wanted to objectively determine what the real sweet spot is for balancing resolution and compression artifacts. I used AviSynth's Compare (http://www.avisynth.org/index.php?page=Compare) filter to measure how much each video varying resolutions deviated from the original. Interestingly enough, the results I got showed that the more resolution (despite the more artifacting) was closer to the original. However, I find this difficult to accept. It is my personal opinion, and the recommended/default action of various automatic video encoders, that retaining maximum resolution for ~ 2hr video at 700MB is almost never the best choice.

So I'm wondering if anyone could confirm these results to see if I made a mistake somewhere. Or, maybe, is compare not the proper tool to use for this situation?

[mf]

You made the mistake of accepting public opinion bullsh*t :). You can go pretty high with resolutions. While I'm at it, forget about compressibility tests ;).

[Jeff Mott]

You made the mistake of accepting public opinion bullsh*tAnd you made the mistake of missing the part of my post that said it is also my personal opinion. Meaning regardless of what public opinion is, I look at various video clips and *I believe* that lower resolutions look better for reducing artifacting.

[mf]

Try Blur(1) next time, you won't need to resize ;).

[Jeff Mott]

Why would you consider blurring better than reducing the resolution? Both will reduce the amount of retained detail, except you can control how much you want to lose more precisely with resizing.

Either way, nothing of what you have said yet has addressed the original question.

[SoonUDie]

I've never done a 1CD encode, so I can't really help you there. But I would say that the best way to decide would be through testing.

In general, though, certain artifacts change their amount of annoyance at different sizes. Take blocks, for instance: the smaller your resolution, the larger the block is in relation to the total area of the frame. If you record to a low bitrate, that means a larger area could end up compressed to the point of looking really shitty.

Also, make sure to test with and without deblocking on in the decoder. A combination of settings which produce more blocking but also more detail may be more pleasing after deblocking than a lower-detail, lower-blocked version.

Personally, I choose my resolution based on destination display size, source quality, and on how important I think visual quality is to experiencing that particular movie.

[mf]

What I mean is that blurring does virtually the same as reducing resolution in terms of "visual pleasingness to the eye". As bits per pixel increase, relative quality deviation lowers. Same goes with picture complexity, the blurrier, the more coefficients can be dropped without creating too much artifacts. The reason I picked the comparison is because people lower their resolution with the sharpest lanczos in the hope that it will still look good, while they're actually just dropping sharpness altogether by lowering the resolution in the first place, while blurring would never cross their mind even an instant, because they'd be destroying sharpness. HELLO, you are already destroying sharpness by lowering the resolution, and actually in a much more crude way than blurring, how crude blurring might seem.

Point of all the above: lowering the res looks more pleasing because you're destroying the picture. Logic of it: if you've already destroyed it, the codec won't have to anymore.
Why keeping the res high actually produces something that's closer to the original: because the codec would never destroy the picture as extreme as lowering the resolution (hell, even RV9 doesn't do that, go figure!).

With all that said, Compare() actually isn't all that good in measuring quality, as it's objective. It's the simple mathematical deviation from the original. Try SSIM for something that's closer to what you percieve.

Why did I start bashing lowering the resolution in the first place? Because I said it's public opinion bullsh*t, and you said it's your own personal opinion. While in fact I think you think that's your opinion, because of the reasons laid down in the explanation above. I could still be wrong though, but at least I know of myself, that I like bashing things that I don't like (wanna get me started on compressibility testing? :D). So there's also some personal satisfaction involved in the venting of frustrations. So don't say my post was useless :p.

[Mug Funky]

hmm... optimum res, huh?

i usually work in constant quantizer with an "acceptable" frame size, usually 384 lines, and witdh depending on AR.

so i'd do 16:9 at 688x384, and 4:3 at 512x384.

i'm never too bothered about file size, as i'm usually encoding half-hour animes, not full movies (i don't have much need of backing up my DVDs except to use them as fodder for my scripting habits).

of the full movies i have done, i usually go 640x480 xvid for 4:3. i fit them pretty well on a CD - no blocks to speak of, and usually no dancing noise (though that's because i filter out the high freqs mercilessly :))

with smart enough filtering you'll be amazed how big a movie you can fit on 1 CD.

[Soulhunter]

Higher resolution... :rolleyes:

Could this be also better coz there would be more MV used then ???


Bye

[Jeff Mott]

The reason I picked the comparison is because people lower their resolution with the sharpest lanczos in the hope that it will still look good, while they're actually just dropping sharpness altogether by lowering the resolution in the first place, while blurring would never cross their mind even an instant, because they'd be destroying sharpness. HELLO, you are already destroying sharpness by lowering the resolutionWell I never said I used Lanczos for resizing. And I never said I was overly worried about sacrificing sharpness. None of what you've said pertains to me or my question.

Why keeping the res high actually produces something that's closer to the original: because the codec would never destroy the picture as extreme as lowering the resolutionThe whole point of this thread is to find what the optimum resolution would be so when we resize it is NOT an extreme destruction. And we do this because when reduced by a proper amount, the picture ends up better than it would have been.

While in fact I think you think that's your opinionWell I'm glad you think that, after all I did explicitly tell you that. :rolleyes:

So there's also some personal satisfaction involved in the venting of frustrations. So don't say my post was uselessOk, then your post was useless to everyone except yourself. The next time you want to vent just yell at your monitor that way we don't have to hear it.

[Jeff Mott]

Could this be also better coz there would be more MV used then ???I don't think there would be _more_ vectors, they just would have greater values at higher resolutions... I *think*. However, Quarterpixel will compress better at lower resolutions. In any case, I don't plan on trying to determine how all different kinds of compression methods perform under various circumstances. I just turn them all on, run some tests and see how they come out.

My situation now is that when deciding which tests come out better, I want to base it on as little opinion as possible. So I need some kind of tool that compares two videos (pictures), preferbly with some kind of basis in human perception rather than a simple mathematical difference.

[Soulhunter]

Originally posted by mf
The reason I picked the comparison is because people lower their resolution with the sharpest lanczos in the hope that it will still look good, while they're actually just dropping sharpness altogether by lowering the resolution in the first place, while blurring would never cross their mind even an instant, because they'd be destroying sharpness. HELLO, you are already destroying sharpness by lowering the resolution
Originally posted by Jeff Mott
Well I never said I used Lanczos for resizing. And I never said I was overly worried about sacrificing sharpness. None of what you've said pertains to me or my question.

- He never said you used lanczos !!!

- He never said you would care about sharpnes !!!

- But he told you some good example regarding common faults... :rolleyes:


Bye

[Soulhunter]

Originally posted by Jeff Mott
I don't think there would be _more_ vectors, they just would have greater values at higher resolutions... Well, thought they are grouped at 8x8 blocks... ;)


Bye

[mf]

Originally posted by Jeff Mott
Ok, then your post was useless to everyone except yourself. The next time you want to vent just yell at your monitor that way we don't have to hear it.
Tisk tisk, don't be so stingy now. We love you too.

[AS]

Originally posted by Jeff Mott
I need some kind of tool that compares two videos (pictures), preferbly with some kind of basis in human perception rather than a simple mathematical difference.

Just an idea, how about you resize differently on a number of clips, then resize them all back to a common res?

e.g. something like

orig = source(blah blah)

a = orig.bicubicresize(640,480).bicubicresize(1024,768)
b = orig.bicubicresize(512,384).bicubicresize(1024,768)
c = orig.bicubicresize(600,450).bicubicresize(1024,768)

interleave(orig,a,b,c)

[Jeff Mott]

But he told you some good example regarding common faults... pThe *point* was that I did not ask for examples regarding common faults. Nor do I need them. The post gave no insight into the original question.

Well, thought they are grouped at 8x8 blocks... Yes, you're right. Though I believe motion prediction is done on 16x16 blocks.

how about you resize differently on a number of clips, then resize them all back to a common res?I already did. In fact I had to in order to use AviSynth's compare filter on them. The only part of this whole process that I'm trying to avoid is using my own personal judgement for the comparisons. I'd prefer to use some tool that would calculate the perceived difference of the reduced/encoded clip from a reference clip.

[avih]

@Jeff Mott:

welcome to the forum.

take it easy. i'm sure you've read the forum rules, especially the one about how NOT to post. people try to help you. they responded. thank them, and if you're not satisfied with the reply then clarify your question.

now, imo, your question was not clear enough in the 1st place. you didn't say which 'recommended' method you're referencing. as far as i know i, there's NO recommendation to keep at maximum resolution. you should use a compressibility test to determine the best resolution. and you do that per movie. movies with much action will be harder to compress at full resolution. movies vith very little action can be compressed at higher resolution without much visual degredation in quality.

there's no default best resolution for "a movie on 700M cd". it depends. movie length is also a factor. if you want to play it safe, go for somewhat lower resolution and you'll get less artifacts (though more blurry image).

[DarkNite]

I'd prefer to use some tool that would calculate the perceived difference of the reduced/encoded clip from a reference clip.

If you'd like a tool to give you some numbers to look at representing the "quality" of the end result in comparison to the original (or any two clips iirc) then here's a link (http://perso.wanadoo.fr/reservoir/avisynth.html) for you.

Just keep in mind that while an objective video quality metric might give you a better idea of how much of the original picture was lost (including noise) it can make no claims as to which looks better.

[mf]

Originally posted by Jeff Mott
The *point* was that I did not ask for examples regarding common faults. Nor do I need them.
I'd say you're trapped on the wrong forum. 90% of threads here on doom9 have to do with opinions and other deviations from the original question. As long as we stay inside the subject this is fine. There is no need to shoot down people on it. But you've really made a nice impression so far. Go on and everyone will love you even more.
The post gave no insight into the original question.
Time to re-read your own post, and mine.
Originally posted by Jeff Mott
Or, maybe, is compare not the proper tool to use for this situation?
Originally posted by mf
With all that said, Compare() actually isn't all that good in measuring quality, as it's objective. It's the simple mathematical deviation from the original. Try SSIM for something that's closer to what you percieve.
Oops! I DID answer your question. Well what a surprise. You were too busy being hot-headed to actually read my post.

[manono]

Hi-

avih has already spoken to Jeff Mott, mf. There's no need for you to pour more oil on the fire.

[morsa]

if anyone is interested, try this ( I did it a long time ago and I discovered nice things).
Take an SD video (I mean 720x525 or 768x576) and compress it with any block based codec to a given bitrate.
Next, take the same SD clip and resize it to a higher resolution (e.g. 2048x1536 ) and compress it again to the same bitrate as the first SD clip.
Third and last: Take the resized clip and compress it to a lower bitrate, may be half or quarter of the first one.
Now tell, which one looks better?

Answer in my next post......

[SILICON]

You are think in one wrong way.

The block are for low bits for codec the image. whith a fixed file size you need drop info. Less info = less blocks.
1 - If you reduce the resolution, you drop a lot of info.
2 - If you blur the image, you drop a lot of info.

The metods 1 and 2, drop the info in a all image, in a regular partern.
Instead.

You can used the full resolution, and drop the info whend you less can see (A dark zones, a noise zones, a very similar zones) and don´t drop the info in critical places (the edges).

Use the filters Convolution, by example, and you can a very good image.