Hello!
I have a very detailed point cloud of an electrical substation. This cloud includes a ton of small lines in the conductors and structures and whatnot.
Is there a way to turn this into a mesh somewhat accurately? I've tried following a tutorial however it just comes out an absolute mess.
Attached is a screenshot of the point cloud, and then one of the corresponding mesh. Any help would be appreciated.
Turning a very detailed point cloud into a mesh
Turning a very detailed point cloud into a mesh
- Attachments
-
- Mesh.jpg (798.29 KiB) Viewed 6680 times
-
- PointCloud.jpg (1.79 MiB) Viewed 6680 times
Re: Turning a very detailed point cloud into a mesh
That's a huge challenge indeed. The key is first to have very good normals (all pointing in the right orientation). But on a cable/wire, that's virtually impossible...
I don't know if some users have food a way to do that...
I don't know if some users have food a way to do that...
Daniel, CloudCompare admin
Re: Turning a very detailed point cloud into a mesh
Thanks for the reply. But that's unfortunate to hear. Was just wondering if there were some settings I could tweak during the normals creation that could avoid columns looking like what's attached. But that's probably expecting too much.
- Attachments
-
- Column.png (654.23 KiB) Viewed 6674 times
-
jedfrechette
- Posts: 46
- Joined: Mon Jan 20, 2014 6:31 pm
- Location: Albuquerque, NM
- Contact:
Re: Turning a very detailed point cloud into a mesh
Where did the point cloud come from and how are you generating the normals?
Hopefully it's from a tripod scanner and you have it as structured scans, as that's probably your only hope for getting good enough normals for something like this.
Like, daniel said this kind of structures is going to be really challenging to mesh with something like Poisson Reconstruction and I'm not aware of any other algorithms that are going to be obviously better. In addition to the cables, PSR also has trouble with thin plates that have points on both sides, and those two types of features are mostly what you have.
Your issue with the columns is likely due to this thin plate issue. To get a good mesh from PSR in this situation the distance between scan points must be smaller than the distance between the front and back surfaces of the steel members in your point cloud and PSR's sampling distance must also be smaller than that distance so you will need a very high density. If the PSR sample distance is larger than the distance between the front and back points on the surfaces with opposing normals both will be considered and you'll end up with the kind of blobby or swiss cheese surfaces that your screenshots show.
Once you've got good normals, I would try meshing small sections at a very high density to try and retain as much resolution from the point cloud as possible. However, even if you max out the resolution it still may not be enough to resolve the front and back surfaces of the steel members. In that case, meshing each scan position separately could also help as then you would only have "front" facing points. If either of those processes give you something that is usable you're going to end up with a bunch of small mesh chunks that will need to be combined depending no how you eventually plan to use them.
Hopefully it's from a tripod scanner and you have it as structured scans, as that's probably your only hope for getting good enough normals for something like this.
Like, daniel said this kind of structures is going to be really challenging to mesh with something like Poisson Reconstruction and I'm not aware of any other algorithms that are going to be obviously better. In addition to the cables, PSR also has trouble with thin plates that have points on both sides, and those two types of features are mostly what you have.
Your issue with the columns is likely due to this thin plate issue. To get a good mesh from PSR in this situation the distance between scan points must be smaller than the distance between the front and back surfaces of the steel members in your point cloud and PSR's sampling distance must also be smaller than that distance so you will need a very high density. If the PSR sample distance is larger than the distance between the front and back points on the surfaces with opposing normals both will be considered and you'll end up with the kind of blobby or swiss cheese surfaces that your screenshots show.
Once you've got good normals, I would try meshing small sections at a very high density to try and retain as much resolution from the point cloud as possible. However, even if you max out the resolution it still may not be enough to resolve the front and back surfaces of the steel members. In that case, meshing each scan position separately could also help as then you would only have "front" facing points. If either of those processes give you something that is usable you're going to end up with a bunch of small mesh chunks that will need to be combined depending no how you eventually plan to use them.
Jed
Re: Turning a very detailed point cloud into a mesh
Hi Jed,
Sorry for the delay. I am using a GeoSlam Zeb Horizon scanner and walking around with it. I am also generating the normals using CloudCompare Normals -> Compute function, and I've tried multiple different settings.
I appreciate your detailed response, that makes sense as to why it's having difficulties. Sounds like it's impossible to use this type of scanner on this type of object and get good normals out of it. That's unfortunate, especially since the point cloud itself looks so good.
Sorry for the delay. I am using a GeoSlam Zeb Horizon scanner and walking around with it. I am also generating the normals using CloudCompare Normals -> Compute function, and I've tried multiple different settings.
I appreciate your detailed response, that makes sense as to why it's having difficulties. Sounds like it's impossible to use this type of scanner on this type of object and get good normals out of it. That's unfortunate, especially since the point cloud itself looks so good.
-
jedfrechette
- Posts: 46
- Joined: Mon Jan 20, 2014 6:31 pm
- Location: Albuquerque, NM
- Contact:
Re: Turning a very detailed point cloud into a mesh
I wouldn't say impossible, just challenging ;-) For a mobile scanner like that you would probably also need to consider the scanner's trajectory in order to make sure the orientation of the normals are consistent. In other works, at a given instant in time, the point at position p1 was captured by the scan head at position s1; therefore the angle between p1's normal and the vector between s1 and p1 must be less than 90 degrees.
I'm not familiar enough with the Geoslam to know if they provide the trajectory data to be able to do that sort of calculation, and I don't think CloudCompare has any code that will do it out of the box, but it could be done. Even then, all of the thin features will still be an issue for your particular dataset.
Good luck,
Jed