CloudCompare forum

Dear cloudcompare community!
I'm currently working on my Master's thesis with TLS-data and I would need some opinion on quality of alignment. I'm aligning scans from a rock structure, and aligning with ICP on the original (manually coarsly aligned by rotation and translation) scans did not work because there are only quite few overlapping parts. So i used point-pair picking and got a RMS of 0.0106791 (which seems quite convenient to me) when using 3 point pairs.
After this, i thought of another approach: I segmented both clouds, so only the overlapping parts (2.5 mio points in one cloud, 3 mio clouds in the second one) were used for the ICP. With this, i got an RMS of 0.0581177. My idea was to then use the transformation matrix also on the other segment of the cloud.
My question is now what of the two results is in general more accurate?
Thanks in advance!
Anna

It really depends on the data. The point-pair picking method is generally more robust if the clouds are quite different. And it can be quite accurate if you use a lot of pairs, and if you can clearly recognize the equivalent points on both clouds.

ICP can also handle clouds with differences (especially with the 'Overlap' parameter correctly set, or with the trick you used). It's also the only option if you have smooth surfaces without clearly identifiable common points / features.

It's hard to tell which solution is the most accurate on a general basis. Pay attention also that the RMS is computed on the point pairs only in the first case, while it is computed on a lot more points (the 'random sampling limit') in the case of ICP. However in the second case the 'pairing' is done automatically so you have less control (you have to trust the statistics ;).