The model defines shapes for performing affine transforms and abstract transforms (scijava.roi.types.AffineTransform3D and scijava.roi.types.AbstractTransform3D). The purpose of the abstract transform is to serve as a hook mechanism for implementation-specific transformations to be supported within the model.
All transforms are shapes. The implication is that all transformations on shapes evalulate to the transformed shape, i.e. the transform shape is the transformed shape.
Transforms between pixel space and physical space (using the unit system defined in the image metadata). Provide both transforms.
If shapes can be defined in either space, should any of these transforms be implicit? If so, when are they applied?
Additional transforms required for display? physical to pixels is equivalent to the modelview transformation matrix. Should we additionally take into account projection/perspective/viewport matrices? Or leave further transformation to the implementor, e.g. starting from shapes reduced to meshes, for OpenGL implementations.
Conversion of shapes to masks needs to happen in pixel space?
In the current model, transforms are specified inline in the shape definition. However, it may make sense to have some transforms out of band in the ROI or shape state, such as pixel to physical (and inverse) transforms. This would require a transform representation with a transform ID as one of its data members.