Assimp中mOffsetMatrix功能疑问：空间转换方向存在认知冲突

I totally get why this is confusing—Assimp's original docs and that later pull request seem to contradict each other, which is super frustrating when you're trying to get skinning right. Let's break this down clearly to settle the confusion:

First, let's address the conflict: the original Assimp docs stated that mOffsetMatrix transforms vertices from mesh space to bone space, but the PR you found points out the opposite is true. Here's the real, confirmed behavior (backed by the PR and subsequent fixes in Assimp):

The mOffsetMatrix actually transforms vertices from bone space (bind-pose local space) to mesh space.

To make this tangible, think about how skinning works at bind time: when you rig a mesh to a skeleton, each bone has a specific transform relative to the mesh. The offset matrix represents that bone's transform at bind time—so if you have a vertex positioned relative to the bone's local coordinate system (bone space), multiplying it by mOffsetMatrix will place it correctly in the mesh's overall coordinate system.

Why the mix-up? It was a documentation error in the original Assimp docs. The code always handled the matrix as bone-to-mesh space, but the docs had the direction reversed, which is what that PR fixed.

For practical use in skinning:

• If you need to take a vertex from mesh space into bone space (to apply animated bone transforms), you'll use the inverse of mOffsetMatrix.
• The mOffsetMatrix itself is used to map a bone-space vertex (in the bind pose) to its correct position in mesh space.

To tie it to the standard skinning formula, here's how it fits in:
finalVertexPosition = sum( vertexWeight * (animatedBoneTransform * inverse(mOffsetMatrix) * meshSpaceVertex) )
Here, inverse(mOffsetMatrix) converts the mesh-space vertex to bone space, the animated bone transform moves it to the animated position, and we sum across all influencing bones to get the final skinned vertex.

内容的提问来源于stack exchange，提问作者Manh Nguyen