CloudCompare forum

Hi,

I'm trying to merge a total of 36 LAZ, all made of 20M to 30M points. However, I can't seem to be able to do it with my computer, I run out of memory everytime. I have 16 GB RAM + 128 Gb virtual memory on a NVME SSD.

I'm following this tutorial (my goal is to create a map based on where I live on a car simulator) : https://www.beamng.com/threads/tutorial ... dar.39370/

Is there any more optimized way of merging everything so I can do it on my computer ? Otherwise I dont know how I will be able to do my project.

Thank you
Etyop

There's a theoretical limit of how much points can be in a single cloud (something like 2 billion). But it doesn't seem to be the issue here (as you should only end up with 2 billion points). Have you watched the computer resources to see how far you go before getting the error? (in terms of memory used by CC).

Hi,

Actually I found a way to go-around the problem. The game in which I need to load the mesh only accepts 200k triangles per mesh (and a max total of 1M triangles).
I run CSF seperately on my 36 LAZ files at 0,5 / 1 and then subsample. But here is my new problem

My cloud contain a lot of trees. CSF manages to remove most of the trees shape, but still leaves tiny bumps here and there.
This makes that when I turn my cloud into a mesh, I have A LOT of tiny triangles making those tiny bumps, which totally ruins my need for less than 200k triangles.

Here is an example :
This is my base LAZ


here it is after CSF


My problem is that I really only need to keep the roads and the relief of the terrain, all the vegetation and habitations must be gone. But as you can see on the image, there still is a lot of points on the vegetation part.

I tried running CSF at 0,2 / 0,4 but it takes a lot of time for almost the same result. I cannot try 0,1 /0,2 because of my computer limitations.

In the tutorial I was following at first (the one I talk about in my first post), the guy uses the Classification scalar field, isolates the ground, and splits the cloud (step 9 on the tutorial).
But I can't try it this way as on my cloud everything is in the same classification, everything is on the same blue line at the total far left of the classification graph.

I know you're french like me so I'll talk a bit more specifically now (sorry future readers)

Au cas où, j'utilise les LAZ dispo au téléchargement du projet LiDAR HD de l'IGN. Le truc c'est que j'ai l'impression qu'ils sont tellement précis que ça en devient difficile de séparer les données. J'ai déjà passé beaucoup de soirées à essayer de séparer totalement le sol et la route, et je voudrais vraiment trouver un moyen de le faire.

Merci d'avance

Yes, CSF is not 'perfect' and often leaves some trees.

One thing you could try is to use PoissonRecon instead of Delaunay triangulation (you can control the quality (octree depth) of the reconstruction, and thus the number of output triangles. It can also ignore small bumps or outliers. Don't forget to request the 'Density' SF output to filter the resulting mesh and only keep the necessary triangles (look at the wiki for more directions).

And I guess IGN haven't classified their datasets sadly... At least not in the free version?

Hi Daniel,
Thank you so much for this answer !
PoissonRecon perfectly fits my needs.

For future readers struggling like me to export LiDAR clouds to only ground mesh :

Run CSF at 0,5 / 1
Subsample random 1,900,000 points
PoissonRecon octree 9

There you have your ground + roads only

Have a very nice day