Cloud to Cloud distance computation

[Jérome]

Hello everyone,

I read the article "CHANGE DETECTION ON POINTS CLOUD DATA ACQUIRED WITH A GROUND LASER SCANNER" written by mister Girardeau-Montaut, mister Roux, mister Marc and mister Thibault.
Three pointcloud comparison process are described in this article: Average distance, Best fitting plane orientation and Hausdorff distance.
I understood that the Hausdorff distance is better than the two other methods.
I also read here http://www.cloudcompare.org/doc/wiki/in ... d_Distance that the C2C distance in CloudCompare uses a kind of Hausdorff distance algorithm, but is C2C as efficient as Hausdorff distance?
An other question, is cloud to mesh distance better than cloud to cloud distance? I read in the article that "meshing algorithms applied to so huge and complex data are slow and hazardous"

Sincerely.

[daniel]

In fact the Hausdorff distance described in the article is the 'same' as the C2C algorithm (it's a partial Hausdorff distance in the sense that we skip the last min/max operation).

And it's hard to answer to the second question as it all depends on the mesh quality. The big advantages of using a mesh as reference are that:
- there are generally less holes in meshes compared to clouds (and the C2C distance is very weak to holes in the reference cloud)
- depending on the way you compute your mesh, you may be able to 'smooth' the noise and therefore get a more 'robust' surface representation. This is true with Poisson Reconstruction for instance, but it's totally wrong with Delaunay. And while the first method can smooth the noise, it can also smooth real features...

I think the real answer to this question is the M3C2 algorithm: it's much more robust than the C2C distance, and probably even more than the C2M distance in most cases (apart if you have a perfect mesh of course).