We prefer to work in physical space, as it makes the shape analysis easier. The difficulty is that there is a poor size correspondence between the landmark size in the 3D view and the 2D views(red, yellow, and green slices). This makes placing landmarks difficult as you can not simultaneously observe them in both the 2D and 3D views. It seems logical that the rendered landmark size would be proportional to voxel size and therefore consistent across different views.

A sample data set is attached. As you will see, when the landmarks are the right size in the 3D view, they are too small to be seen in the 2D views. When the size is appropriate in the 2D view they are too large in the 3D view. This make accurate placement difficult.