MantisBT - ITK
Assuming D is a diffusion tensor and R a rotation matrix, diffusion tensors are usually rotated by
rot(D) = R * D * R^T;
but both R * D and D * R^T are not themselves symmetric matrices. For example let D be an identity matrix and R a non-identity matrix. R*D is R which is not symmetric.
If you try to implement a rotation as D.PreMultiply(R).PostMultiply(R.transpose()), you get incorrect results because the return type enforces symmetry.
My opinion is that these methods should be removed and replaced with a single method named something like Rotate. Only premultiplication or only postmultiplication is probably not a useful operation at least for diffusion tensors. Backwards compatibility should not be preserved here because the current API can never give a correct answer for R * D as the return type enforces symmetry.
I have attached a patch which removes the incorrect methods and replaces them with a usable, though not particularly efficient, Rotate method. I also added to the test class a test that rotating a tensor gives a correct result.
| has duplicate | 0010195 | closed | kentwilliams | Pre-multiply and Post-mutliply functions in itkSymmetricSecondRankTensor |
symmetrictensor-rotate.patch (5,848) 2009-02-20 14:56https://public.kitware.com/Bug/file/2071/symmetrictensor-rotate.patch
SymmetricSecondRankTensorWithRotate.patch (2,269) 2009-04-24 13:09https://public.kitware.com/Bug/file/2187/SymmetricSecondRankTensorWithRotate.patch