Star Wars GOUT in HD using super resolution algorithm (* unfinished project *) — Page 21

@ _,,,^..^,,,_,

Infognition site, well, that's old, of course... what's about the PaNup technique? Spend two minutes to register, and five to read it, and let me know what do you think.

I've checked it. It is image registration and merging. What do you want to hear? It's a decent idea but reminds me of SuperResolution. The theory is much more exciting than the reality. But to be honest, I've never tried this one - if you achieve something wonderful with it, please post a comparsion. After all, if you find two *really* different copies, it should work.

About SuperResolution: [...]

Show me it on the GOUT in comparsion with the traditional methods... it's all I ever wanted.

@ DrDre,

The reason I was defensive is, because you claimed you could achieve the same result with sharpening and denoising, without anything to subtantiate that claim.

Are we back on square one?

I've said *most* of it could be achieved; and there was something to substantiate my claim: Team Blu's project that was already present for a comparsion (but any thought out projects with the same goal would do too). It's all there in my very first post.

As for the caps, I quote myself again: "I've only posted a few caps to show that the [illusion of] resolution improvement can be easily there without SR. Once it is accepted, how can you tell where the sharpening/denoising ends, and where the superresolution begins?"

@ zee944

No we're not going to square one. Your opinion is clear. We disagree. That's it. This discussion has taken up enough space on this thread. 

@ everybody else

I will await the judgement of the wise men (women?) with respect to the AR. 

@ Laserdisc Master

We obviously know R2D2 doesn't have jaggies, but an algorithm like super resolution obviously doesn't. So taking the GOUT as a source it tries to enhance details including jaggies. Another filter might assume all edges are smooth, so it will correct the jaggies in R2D2, but also smooth sharp edges that are supposed to be sharp. In the end no filter will be able to perfectly reconstruct the true source material. A combination of methods could certainly do better, but there will always be some gains and some losses. 

Here's a simple example of image upscaling using a noise free source image. This example shows the difference between sharpening an upscaled low resolution single image and applying super resolution to multiple low resolution images with subpixel shifts. For this example I wrote my own simple super resolution script in MATLAB. The image is first downscaled to half the resolution, than upscaled two times to the original image size (using bicubic). 

The reality of sharpening is that it cannot retrieve details below the pixel level. This is where the difference with super resolution begins. Super resolution does retrieve details below the pixel level. 

The first image shows the original image vs the downscaled + upscaled image:

http://screenshotcomparison.com/comparison/125890

Because the stripes on the girl's pants are at the pixel level, the low resolution image cannot resolve this detail, so it is lost. This lost detail cannot be retrieved by sharpening as is obvious from the next comparison:

http://screenshotcomparison.com/comparison/125891

The information about the stripes is lost in a single image, but can be retrieved by combining the information in multiple images. This is the strength of super resolution, as can be seen in the next comparison:

http://screenshotcomparison.com/comparison/125895

So, although there is overlap between sharpening and super resolution at the super pixel level, super resolution is the only way to retrieve details at the subpixel level. As such super resolution is generally more detailed, more accurate at reconstructing shapes, and has less artifacts, than sharpening. Therefore, if you have a super resolution algorithm available, it makes no sense to use sharpening, as you get all the advantages of sharpening with super resolution anyway. 

@ DrDre,

It's carefully chosen sample material to sell supersampling. You do everything in the book to avoid the comparsion with Team Blu's GOUT.

zee944 said:

@ DrDre,

It's carefully chosen sample material to sell supersampling. You do everything in the book to avoid the comparsion with Team Blu's GOUT.

He asked for samples from it to compare, because he doesn't have Team Blu's release. I sent stills to him via PM, because sometimes Team Blu doesn't like their work criticized and I have a lot of respect for them.

No one had any problems in this thread until you came about, full of demands.

@ zee944

There's nothing carefully chosen about it. It's a simple statement of fact that subpixel detail is lost while downscaling an image. It's also a statement of fact that this information cannot be retrieved from that image once it is lost. Super resolution can retrieve this detail. You can choose to ignore it, but the facts remain the same.

Your second statement is also erroneous, since I did post a comparison to Team Blu's GOUT, and I specifically requested for more screenshots for that purpose.

For some reason you seem to want to antagonize posters in this thread, with deliberately aggressive statements. It should come as no surprise that this behaviour is not appreciated. 

I've just stated what happened so far. There's nothing agressive about it. You could have immediately made me silent with a fair comparsion or at least a promise of it.

On the single screencap you've posted the results were quite close to each other.

Here's the exact same method applied to part of a screenshot from the Star Wars blu ray. 

Original vs downscaled + upscaled image

http://screenshotcomparison.com/comparison/125900

Original vs downscaled + upscaled image + sharpen

http://screenshotcomparison.com/comparison/125901

Original vs super resolution

http://screenshotcomparison.com/comparison/125902

Downscaled + upscaled image + sharpen vs super resolution

http://screenshotcomparison.com/comparison/125903

Just like the previous example, super resolution has more detail and less artifacts. 

@ zee944

You didn't state what happened. You said I did everything in the book to avoid a comparison with Team Blu. This statement was false. I've asked for screenshots of Team Blu's upscale well before you appeared on this thread. 

The screencap I posted was close, but the super resolution upscale had more detail. The openings in the stormtroopers mask on the far left and far right cannot be seen in Team Blu's upscale. They can be seen in the super resolution upscale. There may be other reasons to prefer Team Blu's, but detail reconstruction is not one of them (for this screenshot anyway). 

@ DrDre,

We could argue over and over what happened, but I want to cut it off finally. There weren't Team Blu comparisons posted nor promised (except for one image), although materials were easily accessible. That was my point.

The openings of the two stormtroopers mask on the right is indeed visible on your work. And they would probably never appear without SR. So you're right about that. The grill opening on left stormtrooper is there on Team Blu's too, only it's not as sharp. Your result is sharper overall, and if Team Blu would sharpen more his result, it'd come very close in details. There are tiny extra details on your work, but it's quite close anyway.

I even think that on certain other caps there could be more evidence of SR working.

@ zee944

That sounds reasonable. It is my intention to post more comparisons with Team Blu as soon as I get my hands on their upscale. 

How about just leave ours out of this and do your guys own stuff?

@ _,,,^..^,,,_,

I've written many times that SR works to an extent in "daily" use. PanUp works too. My point was that, on a commonplace material like GOUT it has very little edge over a well done denoising and sharpening. And that should be the reference point.

I don't want to repeat everything I've written in my earlier posts, it'd just lengthen the debate more and more. I've addressed everything you said... and actually, I agree with your post.

@ DrDre, 

I'm glad we're starting to more or less agree on something. Great news, can't wait to see the new caps!

"How about just leave ours out of this and do your guys own stuff?"

Exactly.

Let's all simply stop responding to the problematic user and hope that it goes away.

Mackey256 said:

TeamBlu set the standard for GOUT upscales with their phenomenal Star Wars releases.

There is great work going on here and it's only natural to want to compare it to what has been done before.  When doing that why not compare it to what the community has decided is the best looking upscale done to date?

 Agreed. In no way is the intention to compete with TeamBlu's work, merely to supplement it as an alternative upscale, which others may judge for themselves.

The goal is to find the potential strengths of Super Resolution, and engage in community collaboration to improve upon the strengths of Super Resolution, and debate on what aspects work, or dont work. 

It was the intention of this thread to find out if super resolution can squeeze more detail out of the GOUT. It was never the intention to do a full blown restoration like Team Blu has done or to compete with their great work. Of course as time goes by, and discussion rages on, things can quickly get out of hand. It's understandable that there is a desire for comparison to standards like Team Blu's work. It seems that Team Blu member dark_jedi would prefer no comparison at all. Since this is an open forum, I will follow the majority opinion with respects to this matter, so please state your opinions...

It's very easy to get lost in the weeds when it comes to this stuff.  I tend to lean toward Team Blu's approach of scripting for the best presentation of media that is already starting out sub-par. 

Minimizing the aliasing/jagged pixels is huge to me even if it sacrifices detail on the sub atomic level.  One of your SR examples of the second shot in the film had major aliasing on the star destroyer but Team Blu/Dark Jedi's looks a lot better - maybe this has been addressed in your subsequent versions.

The sweet spot will end up falling somewhere between what's scientifically possible and what's aesthetically pleasing.

Keep up the great work

If you have easy access to either TeamBlu's video, or screenshots, make the comparisons. But if you have to grill somebody for any footage/ screenshots, lets not, as we don't want to upset anyone.

I think the comparisons would greatly benefit the development of the SR script, but if it means ticking somebody off, its not worth it then.

In my opinion this thread isn't fun anymore because people are too touchy and protective, etc.  whatever...  I will come back after the hostility is done.