Landmarks allow us to establish a point-to-point transformation between the source and destination image (without using image intensity features, etc). There is some localization error associated with placing points but this is typically quite small (< 1mm) for landmarks like AC and PC. You need at least 3 points.
The degrees of freedom can vary from one transformation method to the other from rigid registration (6 DOF), to B-splines, and all the way to much higher order nonlinear registrations. The general idea is that the same transform applied to match the points is used to transform the source–>destination with some choice of interpolation method required too. Many of the registration algorithms used nowadays are completely data-driven and do not make use of any anatomical landmarks as priors. However, the other benefit of defining AC and PC is to have an internal reference frame for the location of brain structures – this technique is still commonly used for DBS planning.
I happen to be giving a demo on some work by our team on this very topic at the INCF NeuroInformatics meeting in Montreal this week in case you happen to be around.
Hope this helps,