ICP best registration with 4 points clouds.

[Luca1988SCENE]

Hi, i imported 4 clouds already registered with another Software, and imported as one cloud but CC see them as 4 clouds yet. Now i need to get only one cloud. So i registered the fourth with the third, and the second with the first. Then i registered the two final clouds together and i got a final cloud. i want to know:
-is this method the best one to get the best results with 4 clouds?
-Can i state that CC is more powerfull software than the other to register point clouds?

[daniel]

Well, first it's quite strange that you lose the registration information with an e57 file (if you think that the first software saves the registration information in the e57 file, I would be interested in looking at that file, we might have missed someting in CC ;).

And I would not say that CC is meant to register more than 2 parts together (the registration method are essentially meant to register two clouds before comparing them). Especially because with a method such as the one you used, first it's a bit painful and second it's not the most accurate way to do it (there's no global optimization of the error and you can get global distortion issues due to the errors transmitted from one scan to the other).

[Luca1988SCENE]

importing the first file in e.57 i didn't miss the registration, as you can see in the second photo the cloud is well registered but when you click on it to select it CC selects only one cloud. if you click in a determined part CC selects a cloud (for example XYZ004) if you click in a different position CC selects another cloud ( for examle the XYZ002 one). But i need to work with ONLY ONE cloud because i have to look for the only fit plane. So i thought that it might be better to do the ICP with CC. I did it and the results seems good.i did it with 1.000e(-9) with 1 000 000 points, and with enable farthest points removal. Now CC recognize ONLY ONE cloud that it is what i need.
if you want more info about my work i am available.

[daniel]

Oh then you just have to select all clouds (CTRL key pressed) and then call 'Edit > Merge'.

And indeed, if the scans are already well registered, the ICP registration should perform well.

[Luca1988SCENE]

i have alredy done this before but i thought that through the ICP registration i would obtain better results. In fact with the command "merge" the noise points don't disappear, and with the ICP, ticking the comand "enable points removal" yes.

[daniel]

Oh ok. So it must have indeed refined the registration (as no points are removed). So keep this workflow ;)

[freemountain]

Me again after some time trying out CC.
When I use the ICP registration (e.g. 80 iterations, mesh vs. cloud) what happens when the resolution of the mesh is by far than the one of the cloud? How does the algorithm know which point he avoids? And, does the ICP aim at the polgyon center or rather the vertices to measure the mesh to cloud distance and to do the ICP?
Did you write the ICP procedures yourself or is there in adequate article where you got them from which can be quoted?

[freemountain]

Did I write something wrong? ;)

[daniel]

In fact points are automatically and randomly sampled on the (reference) mesh (at each step so as to compensate for potential "bad luck"). You can increase the number of sampled point with the 'Random sampling limit' spinbox.

And for ICP, the base article is Besl's: http://eecs.vanderbilt.edu/courses/cs35 ... ay_ICP.pdf

[Luca1988SCENE]

hi daniel, there is another problem with the ICP registration. Maybe a bug. When the ICP ends, the new cloud generated (xxx.registered) has only the number of points of the DATA one. how can it possible ?all the points of the MODEL clud disappeared. i thought to get a cloud with a number of points about equal to the sum of the DATA and the MODEL. Can you kindly explain me this result ?

Thanks