CloudCompare forum

Now I'm using the ICP function (ccRegistrationTools).

To optimize my program and produce the best possible results of ICP ,

I want to know more about the setting of parameter of this function.

Is there any document introduce the detail of the parameters?

One more question,

When I use "CONVERGENCE_TYPE::MAX_ERROR_CONVERGENCE", when will the process stop?

"CONVERGENCE_TYPE::MAX_ERROR_CONVERGENCE" mode process how many times is based on what?



Is there a value I can set to stop the process?

just like when finalError smaller than what I set, then ICP end.

Thanks!

Have you looked at the real ICP tool documentation? (http://www.cloudcompare.org/doc/wiki/in ... ?title=ICP).

The 'max error convergence' parameter is the best one. The process ends when the difference of registration error between two iterations falls falls below the specified threshold. This means that the algorithm can't find a much better solution, i.e. it has almost converged.

Specifying the target error is not possible as you don't know what is the minimum error (it's not necessarily 0).

Oh, thanks Daniel.

Sorry, I misunderstood the meaning of the 'max error convergence' parameter.

Hello,
I have a specific question about ICP parameters, how does final overlap influence the final RMS?
I have two clouds of a building, aerial and terrestrial, but they are quite different because of vegetation. I tried with 10% of final overlap and obtained a final RMS of about 0.018 m, but when I create a section there is a shift of about 10 cm between the two clouds. Meanwhile if I set 50% of final overlap, the final RMS increases.
Can you explain me something more about how the final overlap works?
Thank you!

The RMS is computed on the remaining points. If you use less points (and only the best ones) then there's more chance that they fit well. And therefore there are good chances that the RMS remains low.

If you had more points, then you increase the number of potential mismatch. And the RMS is likely to increase.

HOWEVER, this does not mean that the registration is better. Because points can match simply because they lie both on a flat surface in each cloud, but the overal registration can be wrong. So it's generally better to get the most realistic overlap so as to get the most robust registration, but not too many points that don't have equivalents in the reference cloud, as they would shift the final registration.