Skinning formula

Hello,

I’m trying to use skinning with collada and OSG. That’s my first skinning
experience.

The formula is given p 4-8 as:

``````outv = \Sum_i { ((v * BSM) * IBMi * JMi) * JW }
``````

However v is in R^3 and BSM in R^{4x4} so the matrix product is not valid.

Am I missing something? What is the full formula?

I would have expected v to be replaced by [v(0) v(1) v(2) 1] but from my
tests with OSG, it seems more like [v(0) v(1) v(2) 0]. As a consequence
one cannot move the vertices located at [0 0 0].

Am I right? Is there a reason for this choice?

I’m trying to display forces (vectors in R^3) as animated arrows. My plan
is to attach a node to the tip of the arrow, and to translate it according
to the vector value.
In order to do this, I need to discard the orientation of the node, thus the
need for [0,0,0]-located vertices.

Here is a working (in OSG) example, illustrating this idea. The arrow is
a simple 2-vertices line.

http://gist.github.com/610049#file_skin_transl.dae

Here is a non-working one, where the vertex at [0,0,0] does not move.

http://gist.github.com/610049#file_skin … orking.dae

Maybe it would be simpler to scale and rotate the arrow. How would you do that?
Using matrix and scale? Or lookat and scale? Or rotate and scale (what happens
when the rotation axis is [0 0 0])?

Thank you for any insight.

quick hints…

• COLLADA spec generally describes column-major matrix math. This description appears to be row-major (a spec bug).
• Difference between translating a point and a vector i.e. affine vs linear transformation.

Hi Marcus, thank you for you time, again…

I’m not sure to understand what you mean. I believed row/column-majorness was only a matter of storing matrices in linear memory (ie. a matter of software implementation), and that matrix math always (both in French and English speaking countries) used A(i, j) to identify the element at i-iest row and j-iest column. I am right?

Maybe what you mean is that in this formula, v is a row-vector (in R^{1x3}) and not a column-vector (in R^{3x1}) as in the rest of the document. Am I right also?

Here is what I understood on this matter.

Let H = [R p ; 0 0 0 1] in R^{4x4}, with R in R^{3x3} (or SO(3)),

Mapping a point v through H gives Rv +p.
Mapping a vector v through H gives R
v. There is no translation. I don’t know what you mean with “translating a vector”.

Point and vector mapping can be expressed in a way that is consistent with matrix math with homogenous coordinates:

• the homogeneous coordinates of a point v are [v(0) v(1) v(2) 1]
• the homogeneous coordinates of a (“free”) vector v are [v(0) v(1) v(2) 0]

Here is how I understand the problem so far:

• the raw skinning formula exhibits invalid matrix math
• this formula could be valid if one was using homogeneous coordinates (which is not explicitly stated)
• when using homogeneous coordinates, the spec does not specify if the skinned vertices are points or vectors.

Do you agree with this status?

Are they point or vector?

Right and that equation is therefore the transpose of one using a column-vector.

Right. I meant a homogeneous coordinate in a linear transformation where translation occurs.

ok, on this matter I follow you. Sorry for being so picky. I think my loose english does not help here.

[quote=“marcus”]

Right. I meant a homogeneous coordinate in a linear transformation where translation occurs.[/quote]

Sorry, I don’t understand this new sentence.

You haven’t replied to this part. I still don’t know if vertices are points (in which case OSG implementation is probably erroneous) or vectors. Or is the decision unspecified (viewer-dependant?).

I think the formula could be clarified by using the homogeneous coordinates explicitly.

I’m just restating what you said in the context of a translation that is only affine [x,y,z,0] vs linear [x,y,z,1] transformation. I think you understand it.

I think the answer is that it can be either depending on the use-case and semantics. For example, skinning vertex positions will use homogeneous coordinates of the form [x,y,z,1] while skinning normals and tangents will expand to [i,j,k,0]. In other words, it’s a distinction between a position vector and a direction vector.