Jump to content

  • Log In with Google      Sign In   
  • Create Account

Skeletal Animation Weirdness


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
6 replies to this topic

#1 defcanuck   Members   -  Reputation: 103

Like
0Likes
Like

Posted 07 August 2014 - 04:56 PM

I'm trying to debug an issue I'm having with regards to skeletal animation exporting from Maya and importing into my engine - C++ with OpenGL ES 2.0 and GLSL

I export the geometry along with bind pose matrices from an exporter in Maya along with bone/weight information in World Space.  I'm able to see that this data gets imported correctly into my engine.

Maya

http://imgur.com/E21PYkv

 

E21PYkv.png

 

My Engine
http://imgur.com/yUL3D3D

yUL3D3D.png?1

 

As you can see, the base pose transforms are correctly exported.  As is the geometry.  

Next, I export the Animation Data in World Space from Maya.  I'm able to draw the animated skeleton and have verified this as correct.  Furthermore, I've written a custom shader that draws bone indices and weights directly on the mesh - this, too, I've verified as correct.

 

Where things are going wrong is when I pass along the transformation matrix to my vertex shader to displace the vertex positions and normals.  My calculations are as follows - 

 

worldToBindPose - transform from world origin to Bind Pose

worldToAnimation - transform from world origin to Animation Pose

 

finalTransform - the array of mat4 transforms passed to the vertex shader

 

So for each bone in the animation, I perform this calculation (Bind => World => Animation)

 

finalTransform = worldToAnimation * inverse(worldToBindPose);

 

and then in the vertex shader

 

 

vec4 tpos = vec4(pos, 1.0);

    vec4 new_position =

        ((animation[int(bones.x)] * tpos) * weights.x) +

        ((animation[int(bones.y)] * tpos) * weights.y) +

        ((animation[int(bones.z)] * tpos) * weights.z) +

        ((animation[int(bones.w)] * tpos) * weights.w);

    

    return projection * modelview * new_position.xyz;

 

And we get this:

62xGGVx.png?1

http://imgur.com/62xGGVx

 

The previous image sums up the result - the orange-red line is the bind pose, the purple line is the animated skeleton transforms.  As you can see, the geometry should be fit to the purple line, but it's off.  Similar animations display the same result - it looks kind of correct - it's almost as if my matrices aren't concatenating properly.  I've looked at similar code samples, and even a previous exporter I had written years ago for 3D Studio Max, and my math is darn near identical.  

I'm totally stuck, although I think I've narrowed it down to the line

 

finalTransform = worldToAnimation * inverse(worldToBindPose);

 

Any ideas?  

 

 



Sponsor:

#2 Buckeye   Crossbones+   -  Reputation: 4889

Like
1Likes
Like

Posted 07 August 2014 - 05:36 PM

IF your vertices are in world space to begin with, you should be transforming them (assuming you've created the matrices correctly):

 

worldToBindPose * inverse(worldToAnimation)

 

The concept for skinned animation is to move the vertices to bone space, animate them with respect to the influence bone, then move them to world space.

 

Other things to check: are your weights normalized (weight1 + weight2 + weight3 + weight4 = 1, and no weight < 0, and no weight >1)?


Edited by Buckeye, 07 August 2014 - 05:44 PM.

Please don't PM me with questions. Post them in the forums for everyone's benefit, and I can embarrass myself publicly.


#3 defcanuck   Members   -  Reputation: 103

Like
0Likes
Like

Posted 07 August 2014 - 05:53 PM

I throw an error in the Maya exporter if a vertex has a non-normalized weight - so that'd be something I'd catch.  

Vertices are in world space, as are all transforms.  (World to Bind Pose, World to Animation) - I export quaternions for rotation and vector3 for translation, although I did an experiment with exporting raw matrices and turning interpolation off, although this yielded roughly the same result.  

What is of particular interest is I'm able to draw the animated skeleton transforms by transforming vec3(0,0,0) by the current animation transform, then drawing them out manually, and that looks correct.  

I am also exporting the mesh in the bind position.  



#4 Buckeye   Crossbones+   -  Reputation: 4889

Like
0Likes
Like

Posted 07 August 2014 - 08:20 PM


Vertices are in world space, as are all transforms.

 

Vertices being in world space is okay, but a transform isn't in any space.

 

In any case, what happened when you tried what I suggested?


Please don't PM me with questions. Post them in the forums for everyone's benefit, and I can embarrass myself publicly.


#5 defcanuck   Members   -  Reputation: 103

Like
0Likes
Like

Posted 08 August 2014 - 07:03 PM

It still looked incorrect.  But if your vertices are in the bind pose, wouldn't you want to concatenate using the invBindPose?  You're right - you're transforming the vertices from the bind pose to bone space by multiplying the influence verts by the weighted inverse bind transformations, then transforming them using the animation transform (non-inverted).  Sorry if my terminology isn't 100% correct.  

What I'm having trouble grasping is the fact that I am able to draw the skeleton accurately, ie for each bone, I start with a vector at 0,0,0 and then multiply it against that bone's interpolated position, and it looks 1-1 to Maya.  Ditto for the bind pose transformations.  The only issue I see is the skinned mesh isn't lining up 1-1 with the skeleton.

I've attached a .mov file that illustrates the problem.  

Attached Files

  • Attached File  test.mov   513.55KB   14 downloads


#6 defcanuck   Members   -  Reputation: 103

Like
0Likes
Like

Posted 08 August 2014 - 07:06 PM

Also should point out that I've implemented both GPU and CPU animation, and I get the same result from both.  

I've also verified my matrix inverse function is correct (to the best of my knowledge) by taking a sampling of 4x4 matrices, inverting them, and multiplying against the original,to which I get the identity for each one - weird.  



#7 Buckeye   Crossbones+   -  Reputation: 4889

Like
0Likes
Like

Posted 08 August 2014 - 08:16 PM


wouldn't you want to concatenate using the invBindPose?

 

Exactly. The bind pose matrix transforms from bone-space to root-space. The inverse of that is root-space to bone-space, what I assume you mean by "worldToBindPose" (more properly "rootToBoneSpace.") If a vertex is in root-space to start with, you transform it from root-space to bone-space using the inverse-bind-pose. If you have a transform that moves vertices from "worldToAnimation(space)" then the inverse would be "animationSpaceToWorld."

 

Also, since you're using OpenGL, doesn't OGL use right-to-left order for matrix multiplication? I.e., to move a vertex from world-to-bind-space, followed by animation-space-to-world, the final transform would be animMatrix * invBindPose?

 

 

 


I start with a vector at 0,0,0 and then multiply it against that bone's interpolated position

 

What is an "interpolated" position?


Edited by Buckeye, 08 August 2014 - 08:20 PM.

Please don't PM me with questions. Post them in the forums for everyone's benefit, and I can embarrass myself publicly.





Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS