Re: tangent maps

["Ben Rogall" <brogall(at)excite.com>]

Andre, very very nice explanation. I've been using XSI tangents recently in experiments with tangent space vector displacement:
http://www.xsibase.com/forum/index.php?board=29;action="" />It took me a little while to realize I had to do the [0,1] to [-1,1] interval correction. I'm also curious to see whether having only 8 bits per channel for the tangent vector is enough to avoid jumping though I haven't seen anything obvious yet.

Another issue with tangent space is difficulty when you try to get different 3D applications to work with eachother. Once you have the normal and tangent vectors, there are still several ways to determine your T,B,N space depending on which of these vectors you decide to orthoganalize to the others. This choice has to match between the applications. 

I hadn't even thought about the smoothing issue you mentioned. Is that analogous to the normal automatic discontinuity control?

Also is your tangent display utility publicly available?

Thanks,
Ben


 --- On Thu 01/25, =?ISO-8859-1?Q?Andr=E9_Adam?= < a_adam(at)49games.de > wrote:
> From: =?ISO-8859-1?Q?Andr=E9_Adam?= [mailto: a_adam(at)49games.de]
> To: XSI(at)Softimage.COM
> Date: Thu, 25 Jan 2007 11:20:09 +0100
> Subject: Re: tangent maps
>
> Hi Bernard,
>
> the tangent property in XSI is the so called tangent map, or tangent 
> space. Encoded in there you can find a given texture space's u-direction 
> (the reason why you have to reference a texture projection when creating 
> a tangent property) in 3d space (you could call this the "texture 
> mapped" u-direction) on a per sample basis, so we're talking of vectors 
> here. In absence of a more convenient property Softimage decided to 
> encode the 
> tangents into vertex colors. That means you have a little bit 
> of math to do to revert them back to vector data: (VC-0.5)*2. This 
> formula stretches the RGB vertex color values, which have a 0-1 range, 
> back to a vector's -1 to +1 range.
>
> The reason you need a tangent space when dealing with normal maps is 
> that a "normal map in tangent space" (there is also an object space flavour)
> stores relative information as to where a given normal should be bent, 
> which after all is what bump mapping is all about (bump mapping and 
> normal mappinng are practically the same thing). The lila-blueish color 
> in a tangent space normal map means that the geometry's input normal 
> should not be altered, where cyan, magenta, etc colours indicated a 
> shift in normal orientation.
> This shift, which is a relative bit of information as it does not point 
> to global world coordinates, needs a reference on the object itself 
> from 
> which the relative shift should be calculated - the tangent space. The 
> tangent space offers a predefined vector for each sample point and is 
> directly generated from the texture projection the normal map is mapped 
> with onto the object.
>
> I've written a small utility that creates little pointers from the 
> tangent space right on the object's surface to better show the vectors 
> encoded there. I've attached two screenshots which might clearify that a 
> bit. The first one shows a straight spherical mapping, the second one is 
> twisted. As soon as a texture space shows distortions (or even worse: 
> seams), just like the second one, tangent spaces become tricky; the 
> samples for a single vertex point in different directions, which finally 
> would create artifacts during normal mapping. To cope with this issue 
> the tangent operator in XSI has a smoothing value, which basically 
> middles the different 
> sample's tangent vectors using an angle-based 
> approach.
>
> Since the normal map stores its data relative to the tangent space, 
> generation, as done with the ultimapper for example, therefore always 
> needs to reference it during generation to provide correct results. 
> Ultimapper basically marries referenced object curvature from a high 
> resolution source object with the normals of the low resolution source 
> object and the tangent space of the low resolution object.
>
> The Photoshop plugins and tools provided by graphics hardware vendors 
> don't have this kind of 3d information available when generating normal 
> maps from heightmaps. They assume that the object the generated normal 
> map might get mapped onto shows an ideal tangent space, which usually is 
> nonsense. The more distorted the texture space on the mapped object is, 
> the wronger (is that english?!) the bump mapping results with such a 
> textures 
> are. I'd stay away from these tools for production work.
>
> Hope that makes some sense, cheers!
>
>     -André

---

Join Excite! - http://www.excite.com
The most personalized portal on the Web!