Generate mesh from pointcloud

[ang12]

Hello

I'm trying to generate a mesh from a point cloud to export to Unity, etc.

The steps I am taking are as follows:
1. Load e57 file (I am using the sample dataset PumpNoInvalidPoints.e57 (V1) - XYZ,INT,RGB,STR,MUL - 22.1MB from http://www.libe57.org/data.html)
2. For each scan, I compute the normals (Edit -> Normals -> Compute, Surface approximation: plane, Octree Radius: auto, Use preferred orientation: sensor origin)
3. Select all scans and merge into 1
4. Plugins->Poisson Recon, Check output density as SF, all other options default

The mesh generated is a complete mess. Is there something incorrect about this process? Is there a better method?


Thanks

[daniel]

Ah! Sorry it was misleading and probably silly, but the normal orientation was 'from the sensor origin' in this version... while the useful option is probably more the opposite. I fixed it in the latest 2.12.alpha version (where you can now select the 'Towards sensor origin' option).

Mind also that in this file, the second scan (pump2_2) has no associated sensor. You'll have to handle this one differently from the others.

And last, in the general case, pay attention to the fact that the default radius for normal computation might not always be the right one. Especially if you realize that too many normals appear black when looked from the right orientation, then it means that probably the radius was not large enough. And if they appear very smooth, then on the contrary the radius might be too large.

[ang12]

I've installed the newest alpha version (thanks for the fast turnaround!) and am now using the 'Toward sensor origin' option

Do you have a recommendation for a specific setting for the radius, or is this a per-model thing? Similarly, do you have an example of what level of black/smoothness I should be looking for?

Sorry if that is a silly question, I'm relatively new to this

Thanks

[daniel]

The radius depends a lot on the cloud sampling, and the scales/units obviously.

Ideally you should have almost no black normals. And for the smoothness, it depends on the application (it's less an issue for qPoissonRecon for instance)