M3C2 Parameters

[CFGST]

Dear admins and users,

I'm a new user of CloudCompare. After I red the scientific paper on the M3C2 algorithm, I'm trying to estimate mathematically the scale D and the projection d on my point clouds. For this purpose I've tried to calculate the mean roughness sigma(D), using the tool, on the reference cloud and multiplied it by 25 to estimate the scale D. Moreover, since the density of the reference cloud is almost 4 pts/cm², I thought that d = 0.1 m should be suitable to include a large number of points. Is the calculation method above right?


I've tried to compute the M3C2 distances between two point clouds related to a Cliff oriented along the IV quadrant (X, -Y) with the following settings: Scale= 1.5, projection= 0.2, Max depth = 1.5, Calculation mode= Default, preferred orientation = -Y. In attachement you can find a print screen of the M3C2 results. I noted an incorrect sequence of negative and positive distances (Red and Blue points).
Do you have any idea about the origin of this erroneous effect? Is it related to the orientation of the Normals?

Thanks for your support.

[daniel]

Yep, the normals orientation is a key parameter of M3C2.

What is the global orientation of the cloud(s)? And the scale? (can you show the trihedron and the scale bar?).

And what parameters do you get when clicking on the 'auto' button of M3C2 (as a comparison basis).

[CFGST]

Yon can see the global orientation and the scale in the images attached below. Moreover, I've got the following parameters using the 'auto' button of M3C2: D = 0.038541 m, d = 0.077082 m, Max depth = 1.287666 m.

Thank you again.

[daniel]

Ok, then the preferred orientation for normals seems good. However it may give locally wrong orientations on the most 'bumpy' areas. You can compute the normals beforehand with 'Edit > Normals > Compute'. This way you'll be sure that they are ok (don't forget to check the 'use cloud #1 normal' option in M3C2 then). And looking at your cloud and at the values returned by the 'Auto' button, I wouldn't use a radius greater than 5 or 10 cm (i.e. 0.05 or 0.1).

Then for the projection, I would use also 0.05 or 0.1 at most (just as you found yourself). And for the max depth, 1 or 1.5 m seems more than enough.

Can you update me with the results?

[CFGST]

Our aim is to detect small changes in the range of 0 – 10 cm. We have tried several tests with different scales in the order of 0.03≤D≤0.1 m and 0.04≤d≤0.1 m. Unfortunately, the “red-blue” effect was reduced but didn’t disappear. Doing further tests we have discovered that this problem may be related to an incorrect co-registration between the clouds. Note that the co-registration is based only on the ICP matching. Therefore, we think that the reg error included in the M3C2 plugin is not accurate as well as the matching process. In summary, the scales of the M3C2 used in our tests are suitable.
In the future, we have to improve the matching operation so we can provide you with better results.

Thank you again for your collaboration.

[daniel]

Indeed, the registration step is very important. To avoid this kind of artefact:
- either use a small overlap parameter with the ICP tool (as ICP is meant to register clouds that have the exact same shape - if there are differences it's better to ignore the farthest points)
- or use the point pair based alignment tool (if you are able to pick equivalent points that haven't moved in both clouds)