On page 45 there are some screenshots that basically sum up the algorithm. To pull this off, I’d need to access the positions of the vertices of a skinned mesh after the animation is applied. Using the Mesh interface in LateUpdate(), this doesn’t seem to work.
Am I missing something, or is it an unimplemented feature/bug?
You can’t read skinned vertices back on the CPU because they are never written into a readable area. For performance reasons, the vertices are skinned directly into the dynamic write-only vertex buffer (i.e. to the memory where GPU has a fast read access… but the CPU has super-slow read access).
Thus the data of the skinned mesh (vertices, normals, …) is always the source data, not the skinned result.
Edit: maybe you can use some GPU based approach instead? E.g. Motion Blur using Geometry and Shading Distortion (pdf on ATI site)?