I've been trying to get my head around how to properly set up a skinning animation system using the bone transform data that can be imported using the Open Asset Import Library (http://assimp.sourceforge.net) but always seem to fail, which I admit is starting to get on my nerves.

I've made a good 5+ attempts inspired by this article: http://ogldev.atspace.co.uk/www/tutorial38/tutorial38.html as well as the limited hints that are given in the Assimp documentation. One thing to note is that I'm using DirectX rather than OpenGL, so things in the above tutorial will obviously not match to 100%, which may be where I'm going wrong. I'm adhering to the standard matrix multiplication order of scale * rotation * translation and I'm transposing all of Assimp's imported matrices which the documentation suggests are in a right handed format. Comparing my animation frame matrices to that exposed by the assimp_view application, they do seem to match so the problem most likely lies in the way I append all these local matrices in order to construct the final transform of each bone. Pseudo-code of my approach to that:

XMMATRIX transform = bone.localTransform;
BoneData *pParent = bone.pParent;
while(pParent) {
    transform = pParent.localTransform * transform;
    pParent = pParent.pParent;
}

// The actual bone matrix applied to each vertex, scaled by its weight, is calculated like so:
XMMATRIX mat = globalMeshInverseTransform * transform * bone.offsetMatrix;

The other thing I'm doing differently from all info on this I've been able to find is that I'm doing the vertex transformation on the CPU by updating the vertex buffer rather than doing it in a vertex shader. I do see the upside to the vshader approach, I just want to see that I'm able to do it in this, presumably easier, way before moving on to that though.

So my question is basically; is there anything obvious one needs to do when using DX rather than OGL or when setting vertex data on the CPU rather than in the vertex shader (not likely) that I've missed, or is there some more to-the-point tutorial on how to go about this somewhere (I've searched but haven't been able to find any), preferably using DirectX?

Thanks for any pointers,

Husbjörn

You don't mention any particular problems or symptoms of a problem

There isn't that much to mention, my meshes ends up looking like a pile of (wrongly) deformed polygons when I try to set it to match the first animation frame. I've been unable to find any consistency to further pinpoint the error more than to that they are obviously being wrongly transformed. The vertices do seem to keep within a radius that is approximately 2 - 3 times that of the untransformed mesh (in its default bind pose that is) however.

First, just for clarity, is your intent to implement a left-handed coordinate system?

Yes, I've already got a rendering system set up that's working with left-handed coordinates and now trying to add animation features on top of that. I suppose you're right in that this isn't a strict necessity when using DX but it seems to be what is nearly always used anyway so I basically just went with it.

It's not clear what you mean by "right handed format." At a guess: transposing a matrix doesn't change the coordinate system handedness, if that's your intent. Transposition changes the matrix from row-major to column-major or vice-versa.

Ah yes, I got the wording wrong here, they are transposed into row-major matrices. However upon reviewing the Assimp documentation it suggests that its matrices already are in row-major format, however not transposing them makes my static mesh imports appear backwards. Also not transposing the provided joint / frame offset matrices does neither provide the expected results.

Also, as it appears that you're implementing left-to-right matrix multiplication order (SRT vs TRS), then you need to swap the order of parent-child matrix mults

Hm, that does sound like something that may be wrong indeed. I'll take a look at it when I get home.

Also that's a great article you wrote, I've only skimmed through the first half of it so far but it does seem to be rather descriptive, so thanks a lot for that!

Taking a fresh look at your post this morning, the algorithm for the final bone transforms doesn't appear efficient, and, depending on what you're actual code looks like, may be incorrect. You're working your way from child-to-parent. You're duplicating lots of calcs for any bone that has multiple children. You should start with the root frame (node, bone) and work your way through the children - i.e., parent-to-children. That's discussed in the article I linked.

Yes, I'm aware it isn't efficient, its just written to be as obvious as possible so that I don't miss any errors in storing things in flattened hierarchy lists etc. as I intend to do once I get something up and running that I can see what's what from.

Anyway, thanks to your article and some more restructuring into a proper node hierarchy I've finally got the meshes to appear correct, albeit they seem to be offset by 90° around the X axis. I can just offset the root node to fix this but I'm thinking there must be something wrong with my transformations still (I haven't checked but there might be a translation / scale offset too by the looks of it). Anything that sounds familiar?

This is the code I'm using at the moment for applying transformations to the vertices:

            for(UINT m = 0; m < meshList.size(); m++) {
				Mesh *mesh = meshList[m].operator->();
				mesh->LockVertexData();
				size_t vertOffsetPos = mesh->GetVertexDataSemanticOffset("POSITION", 0);
				for(size_t v = 0; v < mesh->vertexCount; v++) {
					XMVECTOR vertexPos	= XMLoadFloat3(&(mesh->pBindPose->at(v)));
					XMFLOAT3 newPos;
					XMMATRIX matTransform	=	flattenedNodeList[mesh->pVertexWeight->at(v).boneId[0]]->matOffset * 
												flattenedNodeList[mesh->pVertexWeight->at(v).boneId[0]]->matGlobalTransform * 
												mesh->pVertexWeight->at(v).weight[0];

					for(UINT w = 1; w < 4; w++) {
						if(mesh->pVertexWeight->at(v).boneId[w] == 0xffffffff)
							break;
						matTransform += flattenedNodeList[mesh->pVertexWeight->at(v).boneId[w]]->matOffset * flattenedNodeList[mesh->pVertexWeight->at(v).boneId[w]]->matGlobalTransform * mesh->pVertexWeight->at(v).weight[w];
					}

					XMStoreFloat3(&newPos, XMVector3Transform(vertexPos, matTransform));
					mesh->SetVertexData(v, vertOffsetPos, newPos);
				}
				mesh->UnlockVertexData(); 
			} 

Edit: I tried to get the formatting to look better for the code section but it doesn't seem willing to play along so sorry for it looking a bit messed up.

Edit2: Ah, figured it out; I forgot to apply the mesh's inverse global transform; everything seems to be working as intended now

Onwards to making a more efficient implementation then, thanks again for your help Buckeye.