I could get perfect registration results (0.025mm RMS error) with basic methods:
- Crop the part of the rod that is not visible in the bone surface scan image (I used MeshLab for this for now)
-
Do a very quick, approximate registration on 3 points using Fiducial registration wizard, harden the transform
-
Use Model registration to fine tune the alignment
Good: All computations are instant (takes fraction of a second). Alignment is perfect.
Not so good: Manual cutting of non-common surfaces and manual prealignment is needed. It takes just 1-2 minutes, but requires attention.
There is a good chance that more sophisticated 3D surface template matching methods would work robustly without these manual steps (see for example OpenCV surface matching algorithm).

