Is there any real value in allowing the user to template over the scalar type? Perhaps it should always be double, unless there's a compatibility problem with the Point class.
Derive this class from TransformBase class. Currently, this class derives from AffineTransform, although it is not an affine transform.
Add a figure in the documentation that informs the formulas used in this class that are used to transform Cartesian to azimuth-elevation-radius
Support vector images
Better support for regions with negative indices.