Have you discovered the amazing IQ of Supersampling AA (ATI 5000 series)

[Tempered81]

MSAA

SSAA (line on the pole is gone, and stitching is way softer)

I never meant for this to be an ATI vs nvidia debate, although it's obvious that the former is doing things a lot better with the HD5000, while the latter has deprecated the system. I am simply trying to explain to you that SSAA is not the best method to render an image, because you always have to interpolate to obtain the final image, and this interpolation can cause a lot of artifacts, the most common is lack of detail but it can get worse, like a moire pattern. And on top of that, it's insanely taxing to run. That's why it's a deprecated technique, and it will completely be forgotten once we have a way to anti alias shader effects.

My turn to thank you for listening.

The shadows look aliased on the MSAA & nice & smooth on the SSAA

 

[Deathray2K]

JAG87, you're confusing subsampling with supersampling, plus an extra dose of misunderstanding what interpolation is.
Subsampling is where a larger image is scaled down. It does indeed need to go through a low-pass (blur) filter to avoid aliasing, which seems to be overdone in nVidia's implementation.
What ATi is doing (As far as I know) is supersampling, which is where each pixel has several samples taken, which are then averaged to produce the final result.
Neither, though, involves interpolation of any variety, as interpolation is only for scaling images up.
MSAA and SSAA (should) anti-alias exactly the same way, with the only difference being which pixels are anti-aliased. SSAA obviously does everything, while MSAA only AAs the depth and stencil buffers. Plus alpha-tests with TR MSAA of course.
SSAA can make textures look softer than MSAA, but only because it's packing much more detail into each pixel, giving a more accurate image. This is also why SSAA (With a high enough sample-rate) doesn't shimmer. Shimmer occurs because the data in the scene (the texture) is a higher frequency (more detailed) than the sample rate (pixel), so as the sample passes over the different parts of the data, it changes without seeming like it should, because you couldn't see parts of the data before. SSAA fixes this by averaging the data out for each sample. This is the much more accurate representation. That's probably also why that black band showed up when it wasn't anti-aliased and not the SSAA one, it was likely a problem that was very minute, but given a whole pixel without SSAA. SSAA's samples mostly didn't include it, so it was minimised. If you look closely, you can actually see a slight dark line in the SSAA, but it's blended with the correct colour, so it's nearly invisible.

With all that junk that out of the way, SSAA sure looks great, I wish I had a card powerful enough to manage it!

 

[divertiti]

does anyone know how to enable SSAA on 4870x2?? any registry hacks or modded drivers?

 

[Phil1977]

The shadows look aliased on the MSAA & nice & smooth on the SSAA

Yes shadows are shader output. And only SSAA can work on shader output

@ Deathray2K: Excellent post. Very technical

 

[SonicIce]

Do they still have combined modes like nvidia had on the 6-series, like 8xS was 1x2SS with 4xMS? And whats up with these blurry screenshots? they look worse than the old 2x Quincunx. SS isn't supposed to look that way.

 

[JAG87]

JAG87, you're confusing subsampling with supersampling, plus an extra dose of misunderstanding what interpolation is.
Subsampling is where a larger image is scaled down. It does indeed need to go through a low-pass (blur) filter to avoid aliasing, which seems to be overdone in nVidia's implementation.

LOL, you really should look up the definition of your words before you use them. Subsampling actually means taking a smaller sample of data from a bigger sample of data, it does NOT mean scaling down. A good example of subsampling is scaling an image down using the nearest neighbor algorithm. In that case only a "subsample" of the data is taken into account discarding the other information. This usually results in an ugly pixelated and aliased image. The "low-pass blur" filter you talk about is actually called a sinc filter, and is itself used as an interpolation algorithm.

What ATi is doing (As far as I know) is supersampling, which is where each pixel has several samples taken, which are then averaged to produce the final result. Neither, though, involves interpolation of any variety, as interpolation is only for scaling images up.

ATI's implementation of SSAA is to take four different samples of the same image using a rotated grid, and then interpolate the final image. Interpolation is a technique of constructing new data points from known data points, in this case constructing less pixels from more known pixels. And it can also be the other way around, making more pixels from fewer known pixels. Interpolation is not only for scaling images up, I have no idea where you learned that.

MSAA and SSAA (should) anti-alias exactly the same way, with the only difference being which pixels are anti-aliased. SSAA obviously does everything, while MSAA only AAs the depth and stencil buffers. Plus alpha-tests with TR MSAA of course.SSAA can make textures look softer than MSAA, but only because it's packing much more detail into each pixel, giving a more accurate image.

LOL. SSAA gives softer images because an interpolation algorithm is needed to construct an image with fewer pixels from larger sample. There are different algorithms, some of which yield poor results while some yield great results. Unfortunately the latter are computationally intensive so to maintain a high frame rate you have to make serious compromises. Try searching for terms like bilinear, bicubic, lanczos, sinc, or spline.

This is also why SSAA (With a high enough sample-rate) doesn't shimmer. Shimmer occurs because the data in the scene (the texture) is a higher frequency (more detailed) than the sample rate (pixel), so as the sample passes over the different parts of the data, it changes without seeming like it should, because you couldn't see parts of the data before. SSAA fixes this by averaging the data out for each sample. This is the much more accurate representation. That's probably also why that black band showed up when it wasn't anti-aliased and not the SSAA one, it was likely a problem that was very minute, but given a whole pixel without SSAA. SSAA's samples mostly didn't include it, so it was minimised. If you look closely, you can actually see a slight dark line in the SSAA, but it's blended with the correct colour, so it's nearly invisible.

Actually as I've said before, textures shimmer because of poorly rendered mipmaps. Shimmering happens even when you are using SSAA, the reason you don't notice it is because the image is rendered at a much higher resolution therefore minimizing the use of mipmaps. The only way to truly get rid of texture shimmering is to use anisotropic filtering.

With all that junk that out of the way, SSAA sure looks great, I wish I had a card powerful enough to manage it!

I agree, but it's not perfect.
Surely much better than using MSAA by itself with no AF and TRAA (or adaptive AA). I'd call it a "jack of all trades"

Do they still have combined modes like nvidia had on the 6-series, like 8xS was 1x2SS with 4xMS? And whats up with these blurry screenshots? they look worse than the old 2x Quincunx. SS isn't supposed to look that way.

Yes they do, you can access them with nHancer.
SSAA does look that way when interpolation is poor. ATI seems much more on the ball with the HD5000's SSAA.

 

[HurleyBird]

Actually as I've said before, textures shimmer because of poorly rendered mipmaps. Shimmering happens even when you are using SSAA, the reason you don't notice it is because the image is rendered at a much higher resolution therefore minimizing the use of mipmaps. The only way to truly get rid of texture shimmering is to use anisotropic filtering.

Wow, I can't even keep count of the number of times people have tried to explain this to you. Yes, you are right, textures shimmer with poorly rendered mip-maps. Nobody disputes that! What people are saying is that there are many other types of image shimmer that AF does not fix. Any kind of aliasing produces shimmer as well. The most obvious example is when there is no anti-aliasing applied to groups of transparent textures, like leaves or a chain link fence. Shimmer due to aliasing can obviously be fixed with anti-aliasing. Edges with normal AA, transparent textures when AA has alpha channel detection, but shader aliasing can currently by two methods: programing anti-aliasing into the shader, or by supersampling.

WHEN PEOPLE ARE SAYING THAT SSAA REDUCES SHIMMER, THEY ARE TALKING ABOUT THE SHIMMER THAT IS CREATED BY SHADER ALIASING. AGAIN, THERE ARE DIFFERENT KINDS OF SHIMMER, AND EVERYONE HERE KNOWS THAT POOR MIP-MAPPING IS ONE CAUSE SO YOU CAN STOP BRINGING IT UP ALREADY.

Thank you!

 

[SirPauly]

Shader, specular, shadows, transparent alpha blends and tests, moire, etc., where developers don't take great care, and most of the softness or slight blurring can be raised by adding some negative lod for the right balance for someone.

It's easy for me: Happy about a feature that may be enjoyed for some gamers that may find some enjoyment; Not use it if the performance hit is too much and enjoy other AA abilities, or simply bitch about an actual choice someone has. Have trouble understanding some bitching about actual choices though.

Personally been an advocate of super-sampled to go along with multi-sampling since both made the move to just multi-sampling -- and thankful that both ATI and nVidia now offer these choices for their customers --- both with pros-and-cons. It's wonderful to see discussions like these for me.

 

[Phil1977]

When the video card renders the image on the screen, isn't it really a "digitized" version of what the scene should look like. Meaning ideally there would be an infinite amount of pixels and colours and so on.

So which final render does look closer to the intended original?

You say that SSAA makes things look smoother. Maybe that's closer to the intended original? Smoother = analogue. Detailed = digital ?

Or put another way can we say that MSAA is a "rougher" representation of the intended original scene? And this "rougher" representation can be interpreted as more detailed?

And from what I have gathered SSAA does "averaging". And doesn't "averaging" mean its closer to the original than just taking a single sample? Whereas with MSAA there is no averaging. It just takes a single sample...

And so far every "how it works" article I found states that SSAA is computational much more demanding, but gives the best results.

 

[Deathray2K]

LOL, you really should look up the definition of your words before you use them. Subsampling actually means taking a smaller sample of data from a bigger sample of data, it does NOT mean scaling down. A good example of subsampling is scaling an image down using the nearest neighbor algorithm. In that case only a "subsample" of the data is taken into account discarding the other information. This usually results in an ugly pixelated and aliased image. The "low-pass blur" filter you talk about is actually called a sinc filter, and is itself used as an interpolation algorithm.

Seriously, just look up signal processing, or anti-aliasing, or anything to do with it really, you'll see that you're confused about what the terms mean, and how it all works.
A sinc filter is an ideal low-pass filter that in practice can't be used.

ATI's implementation of SSAA is to take four different samples of the same image using a rotated grid, and then interpolate the final image. Interpolation is a technique of constructing new data points from known data points, in this case constructing less pixels from more known pixels. And it can also be the other way around, making more pixels from fewer known pixels. Interpolation is not only for scaling images up, I have no idea where you learned that.

No, again you're confusing the term. Interpolation, literally, is estimating points of data between the known points. That's why it's used for scaling images up. This is not something that would be useful for scaling images down.
While

LOL. SSAA gives softer images because an interpolation algorithm is needed to construct an image with fewer pixels from larger sample. There are different algorithms, some of which yield poor results while some yield great results. Unfortunately the latter are computationally intensive so to maintain a high frame rate you have to make serious compromises. Try searching for terms like bilinear, bicubic, lanczos, sinc, or spline.

Again, you seem to have some notion of the terms, but you really don't seem to understand what's actually going on...

Actually as I've said before, textures shimmer because of poorly rendered mipmaps. Shimmering happens even when you are using SSAA, the reason you don't notice it is because the image is rendered at a much higher resolution therefore minimizing the use of mipmaps. The only way to truly get rid of texture shimmering is to use anisotropic filtering.

Sure, that type of shimmering will still happen, but as you said yourself, it is greatly minimised by SSAA. In fact, with sufficient sampling, mipmaps aren't needed at all, since they are designed to reduce aliasing, which SSAA takes care of itself.

I have written a couple software rasterizers, so I have a decent understanding of how AA of all varieties works (Though certainly nothing specific about how either company implements it), but if you doubt me, just read up on signal processing.
Here, I'll start you off:
http://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem

Regardless of how you feel about SSAA, it is more accurate than MSAA, and in fact the only difference there should be is in which pixels get supersampled.

 

[JAG87]

Seriously, just look up signal processing, or anti-aliasing, or anything to do with it really, you'll see that you're confused about what the terms mean, and how it all works.
A sinc filter is an ideal low-pass filter that in practice can't be used.

Sinc resampling is used all the time. The Lanczos filter is itself a windowed sinc filter.

No, again you're confusing the term. Interpolation, literally, is estimating points of data between the known points. That's why it's used for scaling images up. This is not something that would be useful for scaling images down.

Interpolation is used when scaling up or down. You interpolate more pixels from fewer known pixels for scaling up, and you interpolate less pixels from more known pixels when scaling down. I think I can explain this to a 4 year old and they would understand.

Again, you seem to have some notion of the terms, but you really don't seem to understand what's actually going on...

Your right, I only took DSP at university, nyquist shannon was one of the first things we learned. I'm a complete noob when it comes to this topic.

Sure, that type of shimmering will still happen, but as you said yourself, it is greatly minimised by SSAA. In fact, with sufficient sampling, mipmaps aren't needed at all, since they are designed to reduce aliasing, which SSAA takes care of itself.

Shader aliasing is practically negligible compared to the type of shimmering I am referring to. Mipmap shimmering is visible even if you stand 5 feet from the monitor. But you are certainly right, SSAA is the only thing that rids of shader aliasing for now.

And I'm not replying to this thread anymore. All there is to say about SSAA has been said.

 

[BoFox]

Sinc resampling is used all the time. The Lanczos filter is itself a windowed sinc filter.



Interpolation is used when scaling up or down. You interpolate more pixels from fewer known pixels for scaling up, and you interpolate less pixels from more known pixels when scaling down. I think I can explain this to a 4 year old and they would understand.



Your right, I only took DSP at university, nyquist shannon was one of the first things we learned. I'm a complete noob when it comes to this topic.



Shader aliasing is practically negligible compared to the type of shimmering I am referring to. Mipmap shimmering is visible even if you stand 5 feet from the monitor. But you are certainly right, SSAA is the only thing that rids of shader aliasing for now.

And I'm not replying to this thread anymore. All there is to say about SSAA has been said.

Well, BFG10K said that using 8xS (2x2 plus 2x MSAA) or 16xS resolves the blurriness problem that is seen with just 2x2. Also, perhaps AF was not being used in that 2x2 screenshot that you showed us?

You are partially wrong about AF being the only thing that works on mipmap textures, since SSAA (when properly implemented) does some degree of AF by itself (say, 2x2 SSAA automatically doubles the AF by 2x). More on that below..

Anyways, shader aliasing is much more apparent in newer games nowadays. Most games based on UE3 engine has horrible aliasing wherever there is strong contrast with HDR lighting on the edges, and cannot be resolved with anything except brute-force SSAA. Specular mapping is another thing, too, with specular aliasing that plagues many of the DX10 games. Also, adaptive/transparency aliasing do not always work on DX10 games that do not request AA on alpha textures like grass blades (Nvidia's upcoming Fermi has a solution for better Transparency AA on DX10/DX11 games).

You act as if the reduction in texture detail is readily apparent. I'm picky on IQ and can not see a difference in texture quality regardless of whether I'm using SSAA or not. What I do see, is that all shader aliasing is gone, giving me a much more eye pleasing image to look at. With Nvidia hardware it is a different story, SSAA on their cards is a blurry mess - something I expect to be rectified with Fermi.

Until another method of eliminating aliasing emerges, SSAA is only way to get the best possible IQ out of a game right now. Also here is a shot of HL2 at 2560x1600 with 8xS:

http://relm.exophase.com/hl2-2010-01-31-12-21-56-40.jpg

The trees int he distance are 'softer' but remain sharp, not overly blurry like with Nvidia's implementation. Here is another shot at 16xS, total overkill but hey:

http://relm.exophase.com/hl2-2010-01-31-12-31-28-82.jpg

x3sphere said that there is not any blurriness when using 8xS or 16xS, as you can see in the pictures. Like with 8xS, 16xS also uses a combination that includes 2x2 SSAA, but with 4x MSAA instead.

I've also been enjoying SSAA for a good while too. Only a couple games gave me the blurriness problem (DX7 and DX8 games like NOLF), but I think that's when I was using either pure 2x2 or 4x4 SSAA when I was playing with Stereo3D glasses.


I have a couple of questions:

According to this article here: http://www.pcgameshardware.com/aid,...al-quality-to-Radeon-HD-5000-series/Practice/

Does that mean that there's absolutely no way to forced more detailed LOD with ATI drivers newer than 9.11? I think there's an adaptive filter being used with the LOD (where the distant textures are not affected by sharper LOD, but the middle textures are affected somehow) with ATI's new automatic implementation of LOD in SSAA mode.

I mean, by looking at the last picture of Fallout 3 that compares the LOD 0.0 vs Automatic "hidden" LOD optimization vs LOD -1.0, the trees in the background are exactly the same for 0.0 LOD and Automatic optimization, but not with -1.0 LOD. I can see that somewhat higher LOD is being used for textures on the ground in the new automatic mode.


When the 5870 was first released, several people reported that ATI's SSAA method was blurry for some games, but not in other games. I noticed that when BFG10K tested the 8x SSAA mode with D3DAFTester using 16x AF, it did not make the colored mipmaps any smaller. (I'm talking about the colored indications to show the bands where AF is being applied). When I use 2x2 SSAA on my 8800GTX, it literally shrinks the 16x AF colored circles to half the size, making it look like as if I'm using 32x AF. Perhaps it is because ATI refused to let the SSAA make the LOD any sharper with their earlier drivers? Would the recent drivers now allow SSAA to "halve" the size of color bands in D3DAFTester, just like when going from 8x AF to 16x AF?

 

[WelshBloke]

Well it does have profiles, but not as cosy as with Nvidia...

What you do is set your settings. Then go to profile manager, type in a profile name and save the profile...

Then change you settings again, save your profile under another file name and rinse and repeat...

To activate your profile right click on the ATI tray icon or do it from within CCC.

TBH I wish Nvidia did it that way as well.

All my games are on steam (about 25 installed at the mo) and only one shows up in the program setting list, two others show up that I dont own and a load of other stuff I dont need to set AA/AF for (firefox, WMC etc)

And you cant right click on the Nvidia tray icon as there isnt one so you have to either set it all globally or change it manually for each game every time you start them up.

 

[Phil1977]

I had Nvidia there is a quick workaround.

In the driver where you see all the profiles, there is a little tick box. It says something like "display games not installed on this computer".

This will show you every game profile Nvidia has. Works with Steam, I have used it.

 

[Piotrsama]

Interesting thread.
I hate shimmering. And I've experienced it, both on my ATI 4850 and my former Nvidia card.

A tip: (For 5xxx owners).
If you still have XP, all your games will play on DX9 mode, so you can use/try SSAA on any game. (in case you have XP and a newer MS OS, it will be easier to try on XP).

 

[BFG10K]

Anyways, shader aliasing is much more apparent in newer games nowadays. Most games based on UE3 engine has horrible aliasing wherever there is strong contrast with HDR lighting on the edges, and cannot be resolved with anything except brute-force SSAA.

Yep. Even the edges MSAA touches can still have aliasing because MSAA can't affect shader or texture aliasing along polygon edges, so they're still jagged to some degree. No amount of AF will fix that, because AF only can only guarantee the texel value coming out of the texture unit is correct, but it can't correct for what the shader might do to it afterward.

Does that mean that there's absolutely no way to forced more detailed LOD with ATI drivers newer than 9.11?

My 5770 isn't in my rig at the moment, but you could always try ATT to see if the LOD slider is making a difference.

When the 5870 was first released, several people reported that ATI's SSAA method was blurry for some games, but not in other games.

The blurring I reported is a driver bug of some kind related to specific games, because I don't believe it's related to the LOD. Even 2xSSAA was really blurry in games like CoD 4; meanwhile Doom 3 had no discernible blurring even with 8xSSAA. The CoD 4 issue is clearly not normal.

I noticed that when BFG10K tested the 8x SSAA mode with D3DAFTester using 16x AF, it did not make the colored mipmaps any smaller.

They'll move only if the LOD is being adjusted, which oversampling automatically does. This also stacks with AF, which makes sharper images. ATi's SSAA will move only if they adjust the LOD, and that should be happening now.

 

[BoFox]

Yep. Even the edges MSAA touches can still have aliasing because MSAA can't affect shader or texture aliasing along polygon edges, so they're still jagged to some degree. No amount of AF will fix that, because AF only can only guarantee the texel value coming out of the texture unit is correct, but it can't correct for what the shader might do to it afterward.


My 5770 isn't in my rig at the moment, but you could always try ATT to see if the LOD slider is making a difference.


The blurring I reported is a driver bug of some kind related to specific games, because I don't believe it's related to the LOD. Even 2xSSAA was really blurry in games like CoD 4; meanwhile Doom 3 had no discernible blurring even with 8xSSAA. The CoD 4 issue is clearly not normal.


They'll move only if the LOD is being adjusted, which oversampling automatically does. This also stacks with AF, which makes sharper images. ATi's SSAA will move only if they adjust the LOD, and that should be happening now.

Thanks BFG10K. The only ATI card I have right now is a 4870 1GB (and an 8800GTX in my other rig). PCGamesHardware made it sound like the LOD could no longer be adjusted with SSAA and newer drivers for the HD5k series, even if trying to force it using ATT (or even DXTweaker).

This is interesting because whenever SSAA was turned on for Nvidia cards, the AF would automatically double. For example, if I did 2x1 SSAA, the colored mipmap circle in D3DAFTester would be squished to an oval, with double AF being implemented along the horizontal plane. I'm thinking that Nvidia forced the LOD to be sharpened along with SSAA even if the "Clamp" setting was in place (which would normally prevent LOD from being any sharper than 0.0, right)?

Although the article by PCGamesHardware was done a couple of months ago, I think it was quite revealing with the changes in IQ. Hopefully there will be another investigation into the SSAA and AF quality of R800 series GPU's, especially if you can do a more in-depth analysis over at Alienbabeltech. Well, at least comparing it against GTX 480 would be damn interesting (assuming that Nvidia will start officially supporting SSAA again)!

 

[gamechld]

Hi all,

I have a 5970 that I am using in an Eyefinity configuration, so my resolution is 5908x1080 (5760x1080 + Bezel Compensation). Does anyone have any recommendations concerning Anti Aliasing and/or other 3D settings when dealing with a resolution that high?

I feel like SSAA is just a pipe dream at that resolution, but I'm never sure where the best performance/IQ ratio is concerning MSAA vs Adaptive MSAA, and Box vs Narrow Tent vs Wide Tent vs Edge Detect, as well as the sample rate on top of all that. Additionally, I am unaware if different rules and strategies apply when the resolution is that high (like if certain settings have diminished efficacy with increasing resolution).

Basically, I'm trying to figure out what settings make the most noticeable improvements with the least amount of performance hit at this high of a resolution.

On a side note, I have found this thread to be very educational thus far, as my understanding of the mechanics behind all of these graphics settings is limited, so thanks for all your work thus far!

 

[JAG87]

Welcome to the forums.

At that resolution, you need to keep anti aliasing filtering down to a minimum if you want to maintain playable frame rates. I would opt for 4x MSAA, which I would then drop down to 2x Temporal AA (if the game supports it), and finally 2x MSAA. Always try to use in-game MSAA settings rather than CCC, except for Temporal AA.

I would also enable performance Adaptive AA in the driver. It will not give you much of a quality improvement, but it doesn't give much of a performance hit either, so use it when you can (remember that it only works in DX9 right now). Same thing with Gamma correction.

As far as other settings go, I like my texture filtering settings to be all at the highest quality, I believe it's called MipMap in CCC, no optimizations, 16x AF. I force those settings globally, the performance hit from higher quality texture filtering is minimal and the quality improvement is huge.

 

[HumblePie]

Jag87, I will make one correction. Interpolation is VERY specific in definition. It means to create more data points from a smaller data set. No matter if you are talking pure math or signals. Interpolation is the exact opposite of downscaling. They are NOT the same thing and the term can not be applied to both scenarios. I hope even a 4 year old would correct you if you tried to explain it to them as you so bluntly try to put it earlier. What you just said with that statement was so utterly wrong you basically said that positive is negative.

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

but I'll print here the definition in that article.

interpolation is a method of constructing new data points within the range of a discrete set of known data points.

In engineering and science one often has a number of data points, as obtained by sampling or experimentation, and tries to construct a function which closely fits those data points. This is called curve fitting or regression analysis. Interpolation is a specific case of curve fitting, in which the function must go exactly through the data points.

There are some nice charts in that wiki article showing graphs with data points. Then various interpolation methods are applied to each set of data points and the resulting line is drawn from those "interpolated" data points.

Again, you CAN NOT call down scaling as interpolation. They are the exact opposite you numb nut. Now as for your other claims on this subject, I have no idea if you are right or not. However, your major faux paux here on this word and it's definition and usage to me has made me at least less likely to believe anything else you wrote is correct.

 

[JAG87]

Jag87, I will make one correction. Interpolation is VERY specific in definition. It means to create more data points from a smaller data set. No matter if you are talking pure math or signals. Interpolation is the exact opposite of downscaling. They are NOT the same thing and the term can not be applied to both scenarios. I hope even a 4 year old would correct you if you tried to explain it to them as you so bluntly try to put it earlier. What you just said with that statement was so utterly wrong you basically said that positive is negative.

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

but I'll print here the definition in that article.



There are some nice charts in that wiki article showing graphs with data points. Then various interpolation methods are applied to each set of data points and the resulting line is drawn from those "interpolated" data points.

Again, you CAN NOT call down scaling as interpolation. They are the exact opposite you numb nut. Now as for your other claims on this subject, I have no idea if you are right or not. However, your major faux paux here on this word and it's definition and usage to me has made me at least less likely to believe anything else you wrote is correct.

You must have a very limited IQ since you can't even read english. Let me write the definition down one more time for you:

Interpolation is a method of constructing new data points within the range of a discrete set of known data points.

Where does it say you have to make more from less? Anytime you create new data points through estimation algorithms, you are interpolating.

And anyway don't bring up this conversation again, I only replied to help the newbie, and I don't care to talk about this crap again with people that have no knowledge. Go take some dsp classes.


EDIT
Here, I'll give you an actual encyclopedia definition of interpolation:

In mathematics, estimation of a value between two known data points.

Estimation of a value between two known points, that's down scaling by definition. And when you upscale, you use new interpolated points together with the original points to generate a larger data set.

 

[HumblePie]

You must have a very limited IQ since you can't even read english. Let me write the definition down one more time for you:

Interpolation is a method of constructing new data points within the range of a discrete set of known data points.

Where does it say you have to make more from less? Anytime you create new data points through estimation algorithms, you are interpolating.

And anyway don't bring up this conversation again, I only replied to help the newbie, and I don't care to talk about this crap again with people that have no knowledge. Go take some dsp classes.


EDIT
Here, I'll give you an actual encyclopedia definition of interpolation:



Estimation of a value between two known points, that's down scaling by definition. And when you upscale, you use new interpolated points together with the original points to generate a larger data set.

Er no. down scaling is removal of data points. Or removal of two data points to create a single new data point. With interpolation you still have all the original data points in the set. You change none of the original data. You only add estimated new data points that would fit in the existing data set.

So there is where you are clearly wrong.

Here is an example

I have the data set

1, 5, 7, 9, 13

I can interpolated the numbers 3, and 11 to fit the set. How I come up with those numbers is based on the formula used. In this case, linear interpolation works perfect since it is a straight line.


For down scaling it is the opposite. I have the following 9 data set but need 5 data points

1, 3, 5, 7, 10, 11, 13, 15, 17

I would downscale it to the following:

1, 5, 9, 13, 17

See what happened here? I removed existing data, AND because the 10 point was a bit off for a straight line, it was adjusted to 9 to fit the set. I estimated the best set from the previous data set with a smaller sub set of data points. Again, DOWN SCALING IS NOT INTERPOLATION. I don't care how much you want to say other wise until you are blue in the face, but you are still wrong. Damn, you remind me of the idiots that keep claiming 0.999... does not equal 1.

And as for the "dictionary" definition, since I guess wikipedia isn't enough for you...

verb (used with object)
1.
to introduce (something additional or extraneous) between other things or parts; interject; interpose; intercalate.
2.
Mathematics. to insert, estimate, or find an intermediate term in (a sequence).
3.
to alter (a text) by the insertion of new matter, esp. deceptively or without authorization.
4.
to insert (new or spurious matter) in this manner.

Again, it is to find INTERMEDIATE term in a sequence. Meaning, you keep the original and add more data points. Notice with each definition it is to ADD something new usually by insertion into something that is already there. You can NOT interpolate by REMOVING. So in this you are absolutely dead wrong and pwned.

 

[JAG87]

Er no. down scaling is removal of data points. Or removal of two data points to create a single new data point. With interpolation you still have all the original data points in the set. You change none of the original data. You only add estimated new data points that would fit in the existing data set.

So there is where you are clearly wrong.

Here is an example

I have the data set

1, 5, 7, 9, 13

I can interpolated the numbers 3, and 11 to fit the set. How I come up with those numbers is based on the formula used. In this case, linear interpolation works perfect since it is a straight line.


For down scaling it is the opposite. I have the following 9 data set but need 5 data points

1, 3, 5, 7, 10, 11, 13, 15, 17

I would downscale it to the following:

1, 5, 9, 13, 17

See what happened here? I removed existing data, AND because the 10 point was a bit off for a straight line, it was adjusted to 9 to fit the set. I estimated the best set from the previous data set with a smaller sub set of data points. Again, DOWN SCALING IS NOT INTERPOLATION. I don't care how much you want to say other wise until you are blue in the face, but you are still wrong. Damn, you remind me of the idiots that keep claiming 0.999... does not equal 1.

And as for the "dictionary" definition, since I guess wikipedia isn't enough for you...



Again, it is to find INTERMEDIATE term in a sequence. Meaning, you keep the original and add more data points.

The part I bolded is called bilinear interpolation. Technically you need a 2x2 sample to apply bilinear interpolation, but I'll make a stretch in this case because of your poor data set choice.

The rest of what you did to that data set is also called interpolation, and you used a nearest neighbor algorithm, where there is no averaging taking place, but simply taking the closest known point to the interpolated point. It is an extremely fast interpolation algorithm which produces poor results.

Seriously dude there's about half a dozen links in this thread about this, read before you post.

Here I'll save you the trouble.
http://www.cambridgeincolour.com/tutorials/image-interpolation.htm

 

[HumblePie]

The part I bolded is called bilinear interpolation. Technically you need a 2x2 sample to apply bilinear interpolation, but I'll make a stretch in this case because of your poor data set choice.

The rest of what you did to that data set is also called interpolation, and you used a nearest neighbor algorithm, where there is no averaging taking place, but simply taking the closest known point to the interpolated point. It is an extremely fast interpolation algorithm which produces poor results.

Seriously dude there's about half a dozen links in this thread about this, read before you post.

Here I'll save you the trouble.
http://www.cambridgeincolour.com/tutorials/image-interpolation.htm

And on that website, even for 2x2, the original data points remain the same. Again, interpolation is for adding new points, not removing old points. Is it possible to do both downscaling and interpolating? Yes. But they are not the same.

Take my data set above where I showed the down scaling. I guess if you change the order in how you did it such as taking the following 9 point set:

1, 3, 5, 7, 10, 11, 13, 15, 17

and making the following 4 point set:

1, 5, 13, 17

and then interpolating the 5th point to make the final set:

1, 5, 9, 13, 17

then you technically did both down scaling and interpolating.

However, you could just apply a single downscaling routine and come up with the same 5 point set without doing interpolating.

This is where you have been confused. By definition, interpolation is ADDING points. The new points can be exact or estimates. But interpolating by itself does not change the existing data points. I am not arguing over the AA crap you were talking about before, but in this instance you are absolutely wrong. Just admit it and get over it.

 

[JAG87]

And on that website, even for 2x2, the original data points remain the same. Again, interpolation is for adding new points, not removing old points. Is it possible to do both downscaling and interpolating? Yes. But they are not the same.

Take my data set above where I showed the down scaling. I guess if you change the order in how you did it such as taking the following 9 point set:

1, 3, 5, 7, 10, 11, 13, 15, 17

and making the following 4 point set:

1, 5, 13, 17

and then interpolating the 5th point to make the final set:

1, 5, 9, 13, 17

then you technically did both down scaling and interpolating.

However, you could just apply a single downscaling routine and come up with the same 5 point set without doing interpolating.

This is where you have been confused. By definition, interpolation is ADDING points. The new points can be exact or estimates. But interpolating by itself does not change the existing data points. I am not arguing over the AA crap you were talking about before, but in this instance you are absolutely wrong. Just admit it and get over it.

What you are talking about is sub sampling, and it is the easiest form of down sampling, which involves no interpolation. The best way to down sample is to interpolate new points.

For your data set:

1, 3, 5, 7, 10, 11, 13, 15, 17

The best 5 point down sampling would be:

3, 6, 9, 12, 15
average difference of 3

This reduces the aliasing that your shitty sub sampling algorithm produces:

1, 5, 9, 13, 17
average difference of 4


Sub sampling and interpolation are not the same thing, in fact they are opposite just like you said.

However good down sampling involves interpolation, you're just a stubborn kid who won't admit he's got his terms mixed up. And with that, I'm done replying to this thread.

Edit
I just realized I called nearest neighbor an interpolation algorithm in my previous posts, and I would like to correct myself, there is no interpolation taking place.