I am trying to use the geometric feature functions in the Cloudcompare and wasn't sure of many features you have there. Roughness, curvature, density sounds straight forward but most of the indices under the 'feature' didn't click. For example, Sum of eigenvalues, Ominvariance, Eigentropy, Anisotropy, Planarity, Lineariy, PCA1 and 2 and surface variation.
How are you all using these indices in measuring surface structure? Can you please explain what those are and when/how they can be used? If possible, can you also suggest a resource (book or website) where I can gain such information? 3D vision is not directly related to my research field, so I am having difficulties understanding those indices. I do know PCA though...
Kwan
Geometric feature indices and what they mean.
-
JamesRayen
- Posts: 19
- Joined: Tue Aug 23, 2016 1:33 am
Re: Geometric feature indices and what they mean.
You can read the paper " Contour detection in unstructed 3D point clouds" as most of the features are defined in the paper
Re: Geometric feature indices and what they mean.
I have put relevant effort to find the meaning of these indices, how they are computed but I still do not get it. I am not a mathematician, but have lots of experience especialy with PCA. The paper suggested by JamesRyen "Contour detection in unstructed 3D point clouds" is really useful, but does not explain how the PCA is conducted (even the term of PCA is missing from it): what are the variables for the PCA, how the local neighborhood radius affects it.
Although it seems a very valuable tool to analyse the point clouds, without a user-friendly explanation it stays unused.
Please help, how these indices are computed and can be interpreted.
Although it seems a very valuable tool to analyse the point clouds, without a user-friendly explanation it stays unused.
Please help, how these indices are computed and can be interpreted.
Re: Geometric feature indices and what they mean.
PCA = Principal Component Analysis.
You take a point cloud (here the neighbors of each point), then extract the "principal components" --> in effect those are the eigenvectors (= directions) attached to the eigenvalues of the cross-covariance matrix. Basically the 2 first ones (attached to the biggest eigvenvalues) are the directions along which the cloud is the most elongated. The 3rd one (attached to the smallest eigenvalue) is the diretion with the list elongation (the 'flat' dimension).
Considering the eigenvalues or eigenvectors gives you interesting information about the shape of this cloud.
If you do this on the set of neighbors around each point, you get some info about the local shape of the point cloud /surface.
And if you change the radius of the neighborhood, the scale at which the 'geomorphologic' info is extracted changes.
You take a point cloud (here the neighbors of each point), then extract the "principal components" --> in effect those are the eigenvectors (= directions) attached to the eigenvalues of the cross-covariance matrix. Basically the 2 first ones (attached to the biggest eigvenvalues) are the directions along which the cloud is the most elongated. The 3rd one (attached to the smallest eigenvalue) is the diretion with the list elongation (the 'flat' dimension).
Considering the eigenvalues or eigenvectors gives you interesting information about the shape of this cloud.
If you do this on the set of neighbors around each point, you get some info about the local shape of the point cloud /surface.
And if you change the radius of the neighborhood, the scale at which the 'geomorphologic' info is extracted changes.
Daniel, CloudCompare admin
Re: Geometric feature indices and what they mean.
Thank you for the quick and informative answer, it helped me a lot.
So, just to be sure about the indices here: the local neighborhood radius I set defines a certain amount of points which is the basis of PCA models and it is repeated for each point with different set of points having different geometric shapes? And the PCA performed here is not a multivariate method but instead tries to find the directions, i.e. the longest axis, second longest? So it is more about eigenvalues than PCA itself (I was stuck with the statistical approach of PCA)?
What can be an optimal value for the local neighborhood radius?
Is this method good for ALS scanned data? I experienced that e.g. curvature or number neighbours depend on the overlapping paths so cannot be applied, but there seemed reasonable e.g. with PCA1 and PCA2 ...
So, just to be sure about the indices here: the local neighborhood radius I set defines a certain amount of points which is the basis of PCA models and it is repeated for each point with different set of points having different geometric shapes? And the PCA performed here is not a multivariate method but instead tries to find the directions, i.e. the longest axis, second longest? So it is more about eigenvalues than PCA itself (I was stuck with the statistical approach of PCA)?
What can be an optimal value for the local neighborhood radius?
Is this method good for ALS scanned data? I experienced that e.g. curvature or number neighbours depend on the overlapping paths so cannot be applied, but there seemed reasonable e.g. with PCA1 and PCA2 ...