Have tried some matrix initialization from the image.
Still initializes shape model upside down for images with non-CT geometry.
As debugging is not possible due to non-stop exceptionsin the "segmentation" module, I am not able and willing to continue working on this bug.
I will not remodel support for non-uniform geometries again.

Hannes, contact me once segmentation module is in a half stable condition, then we can think of a solution.

By then, for 3M3, only the Liver shapemodel will work in any useful way.

Okay, fixed.

[SVN revision 20092]
FIX (#2928): Initialize shape model by image matrix

[SVN revision 20174]
FIX (#2928): Only do default transform and matrix init if no automatic method is available

[SVN revision 20184]
FIX (#2928): Fix liver default init (actually it never worked before...)

[SVN revision 20235]
FIX (#2928): Don't use matrix for LV wall models because of shear strain

For the LV wall models, the geometry matrix is a matrix product of rotation and shear strain. Because the models were trained on the sheared data, they do not need to be sheared again explicitely.
But, sole rotation can't be extracted or estimated from the matrix.

Only possible solution would be (tip from Thiago):
Use matrix on a Cube. It is sheared, so calculate she shear matrix from the cube angles. Apply inverse shear matrix on geometry matrix -> obtain rotation.

Too much effort for a simple shape model init.