Getting weird results from Poisson Recon

[DaanS]

Hi everyone,

This is my first time trying out CloudCompare and I am facing some weird outcomes from the Poisson Recon Plugin/ Algoritm.
I have a Point Cloud segment from a tunnel and I am trying to convert this Point Cloud into a Mesh or Surface.
I saw some videos on YouTube and some forum posts here with promising results, however when i tried to copy those workflows onto my Point Cloud I didn't get the results I was hoping for:

1.PNG (394.01 KiB) Viewed 2309 times

I tried using some new workflows from these topics but to no avvail:
https://www.danielgm.net/cc/forum/viewt ... a68#p24215
https://www.danielgm.net/cc/forum/viewt ... dba06f9457

Would love to hear what I am doing wrong and how to get a desired result.

Kind Regards,
DaanS

[daniel]

So first you should reduce the mesh extents thanks to the 'Density' scalar field (this way it might be easier to understand what happens).

Next, it's hard to see how are your cloud normals... You have to make sure they are in the right orientations.

Can you maybe share the cloud with us? (admin [at] cloudcompare.org).

[DaanS]

Hi Daniel,

Thanks for the quick reply!
I have sent the Point Cloud over e-mail.

Kind Regards,
Daan

[daniel]

Sorry, can you send it again? I believe it was in the spam folder that I flushed without looking at the contents!

[DaanS]

Hi Daniel,

I think I did something wrong when trying to Compute Normals in my initial approach.
I have now found a Workflow that works great and exactly as expected!

01.png (451.11 KiB) Viewed 2242 times

02.png (343.3 KiB) Viewed 2242 times

03.png (343.57 KiB) Viewed 2242 times

04.png (643.2 KiB) Viewed 2242 times

05.png (523.56 KiB) Viewed 2242 times

08 (Octree Depth of 12).png (1.24 MiB) Viewed 2242 times

[daniel]

Thanks for the feedback!

[jakess]

Hi everyone,

Thanks DaanS for the nice explanation! I tried to use the same way, with a few adjustments on my point cloud, where I want to reconstruct tree stems.
To compute the normals I used: the quadratic model, octree 0.1 and knn 6. In the poisson tool window I selected octree depth 8 and in the advanced section: samples per node 3 and point weight 0.

Somehow the poisson algorithm only reconstructed the stems where the normals were black (which I understood was not good?) and ignored the other parts (the colors are a result of the Ransac plugin I used before on the cloud).

I attached an image of the normals and the resulting poisson recon below:

normals.PNG (152.64 KiB) Viewed 2193 times

poisson.PNG (446.49 KiB) Viewed 2193 times

Could the way I computed the normals be the problem for the patchy result?
Any input to improve my result is appreciated.

Best,
jakess

[daniel]

If the normals are black, it may just mean that they are inverted. Have you tried to invert them to check that? (Edit > Normals > Invert)

[jakess]

Hi Daniel,

Yes, I tried to invert them. But it just changed the stem part with the "good" normals to have the black ones and vice versa. For the poisson recon I could not see a difference. I also tried to use the "preferred orientation option" but could not figure out which setting to use as the normals around the trees do not only have one direction but more a preferred plane (like XY-plane).

[daniel]

Anyway, I don't clearly see what the 'foliage' points look like, but you would need quite a lot of points on the leaves (with good normals) to be able to reconstruct them as a mesh.