Help me understand anisotropic filtering better!

[clement]

I've been a LOOOONG time lurker of theese boards(although I was really in and out with all the stuff that happend with this site), and I still think I'm having a hard time figuring what anisotropic filtering actually does. From what I can tell it seems to allow a lower LOD Bias or something (without the negative aspects normally equated?), as that's the only thing I can tell between aniso on and off. I think anisotropic filtering is supposed to eliminate texture aliasing on polys at extreme angles (and odly shaped surfaces, like spheres too right?). Basically trying to figure out what to look for in screenshot comparisons between two cards, the only difference I can see is bluriness off in the distance, is that all there really is to it?

Thanks!

 

[Pete]

I wonder what prompted this thread?

 

[noko]

Start here at Beyond3d Glossary:

http://www.beyond3d.com/words/page_a.php

Come back and let us know if that helps 8).

 

[Doomtrooper]

Another more technical link...

http://www.sgi.com/software/performer/brew/anisotropic.html

 

[Dave Baumann]

This one might help a little as well:

http://www.beyond3d.com/articles/Anisotropic/index1.php

 

[clement]

aha! Tanks! That's what I thought it was, but since I didn't KNOW it was that, every once in a while i'd see a conversation about Aniso. and it would seem they were talking about something I didn't see! Much clearer.

Now I almost feel a little embarrased that I forgot to check the glossory!(how long has it been there anyways?) Who would of thunk it?

I still don't see how some of you guys can have such picky eyes, sometimes still have problems picking out the anisotropy after it is pointed out! But then again I'm still in the stone age as far as my 3d accelrator goes. (Voodoo3 PCI, hey, this was originally in a Pentium Pro machine with no AGP! )

BTW (after about 10 edits...) looks like the old glossory could use a little updating, where are Vertex and Pixel shaders? Don't worry I'm pretty sure I got a good grasp on that.

 

[Bambers]

It can seem picky but you only really notice it once you lose it. If you played fps games (counter strike is particularly noticable for aniso) for a few months with 16x aniso you'd notice it if you went back to normal.

 

[barbarella]

Clement,

if you read french tries that article :

http://www.onversity.com/cgi-bin/pr...&M=informat&O=touslesmots&P=a1201

 

[Gunhead]

Clement, congrats for getting an-iso-tropic spelled right at first (outta-lurking) attempt

Not something everybody has done...

And possibly indicative of a growing interest in, and awareness of image quality! Not that steady framerate didn't count more, but still.

[off topic] And why is it vertex *shaders*? No *shading* is done there, just programmable geometry.
<- bitching to himself [/off topic]

 

[Reverend]

Clement, congrats for getting an-iso-tropic spelled right at first (outta-lurking) attempt...

I've always pronounced it as ani-so-tro-pic even though I pronounce isotropic as i-so-tro-pic.

 

[Simon F]

Clement, congrats for getting an-iso-tropic spelled right at first (outta-lurking) attempt...

I've always pronounced it as ani-so-tro-pic even though I pronounce isotropic as i-so-tro-pic.

???I'm confused. Anisotropic just sounds like "an-isotropic".

To answer the original poster, if you have access to a good library (or a technical bookshop that doesn't mind you doing some reading ) then take a look at
"Computer Graphics: Principles and Practice" (Foley, van Dam...) pages 826-828. It has some nice illustrations of the quality of various texture AA schemes. (BTW ignore the structure of the MIP map. I don't think anyone does it that way anymore).

 

[KimB]

Only thing is, pronouncing it "an-isotropic" just sounds too much like "one isotropic" to me. I generally pronounce the 'i' as a short 'i' in aniso. I pronounce isotropic with a long 'i'.

 

[clement]

Another quicke!

Will anisotropic filtering get rid of temporal aliasing? (What about other types of ailiasing?) Simple yes or no would do here, thanks! (Obviously I can't test this myself, and such is something that needs to be seen in motion)

I'm just gld I didn't spell it Anistropic, as sometimes I leave out that 'o' since it seems to roll off the tounge easier. (Didn't think there would be a whole discussion on this )

 

[KimB]

Anisotropic doesn't remove aliasing any more than normal bilinear/trilinear filtering does. What it does do is allows a higher texture LOD with the exact same amount of aliasing for angled surfaces (When done properly, of course...).

 

[Ailuros]

If you don't mind a minor addition:

iso = even, equal
aniso = uneven, inequal

tropic comes from tropos = manner, way, mean

 

[Simon F]

Anisotropic doesn't remove aliasing any more than normal bilinear/trilinear filtering does. What it does do is allows a higher texture LOD with the exact same amount of aliasing for angled surfaces (When done properly, of course...).

I'd have to disagree here. The problem with isotropic filtering (e.g. bilinear/trilinear MIP mapping) is that the filter kernel is always square which means that, for quite a number of pixels, you have to balance between either over-filtering (i.e. blurring the texture) to guarantee no aliasing, or choose to under-filter which means you will have aliasing.
If we assume that an application has biased the MIP map access so that it choosing somewhere between the two 'extremes' then there will still be some aliasing. Anisotropic (long first 'i' ) filtering should therefore have less aliasing.

 

[KimB]

Yeah, you're right...I actually went to webster.com, and it's definitely a long 'i'

You can always say that a higher-order filtering removes aliasing, but the truth is that no graphics hardware has any excuse in showing any significant texture aliasing with default texture LOD. It is just irresponsible for graphics hardware to set LOD's that will introduce aliasing.

From this perspective, anisotropic should not decrease aliasing, as texture LOD should be adjusted to compensate.

Obviously, if texture LOD was kept identical with aniso enabled or disabled, aniso would certainly improve texture aliasing by a significant amount. It just depends on how you look at it.

 

[Simon F]

Chalnoth:
With MIP mapping and bilinear/trilinear it is possible to guarantee no aliasing but you this means that at times you will overfilter the textures, i.e. end up removing some of the detail/higher frequencies that you actually want to keep.

If you look at the reference I gave in Foley and van Dam, you'll see what I mean. Anistropic filtering improves things in these situations (but at a cost, of course).

 

[KimB]

Well, it does seem theoretically possible, but I do wonder as to what happens when the texture u/v lines are not parallel to up/across on the screen. I know that there isn't any additional noticeable aliasing in these situations on my old GeForce DDR or current GeForce4 Ti 4200, but that doesn't mean it isn't there.