Paramétrage segmentation/ décimation

Added by Domenech Anthony about 3 years ago

Good evening Mr. Hackenberg Jan,

I need your advice.
I’m looking for the best density relation of my point cloud based on the segmentation parameters.

What are the best segmentation parameters for decimated point clouds at 5 cm cube and decimated at 2 cm cube?

Pending your reply I remain at your disposal for any additional information.


Replies (3)

RE: Paramétrage segmentation/ décimation - Added by Hackenberg Jan about 3 years ago

Dear Anthony

as far as I remember you work with a tropical rainforest plot right? Rather large and therefore low resoluted.

Downscaling has to purposes:
- fastening the exectution time
- evening the point density.
The later one is more important, as the first only affects the computation time. What you (and other people) can do is by applying the scanning resolution compute manually the distance between two neighboring points in the maximum distance to the scanner. That is the value which you never should increase as a voxelgrid downscaling size. If this value is rather large, you can make it a bit smaller. You will not have perfect even density cloud but save more points on the more important stems which are closer to the scanner.

Maybe 5cm is under what I wrote a good choice. Do the little computation yourself and test. I know only that I designed for higher resoluted clouds and cannot say how my method would react to 5 cm average distance. I recommend going for 2 cm as I already did modelling with this resolution. I still consider 1 cm better and that is my standard parameter for our scanning design.

I hope this gives you some background. Now my recommondation.
  • Speed of computation = set to 5 and pay with runtime.
  • range = I assume 5cm have been tested, so try out 10cm, then 20cm. The larger the more problematic might became the subdivision in dense undergrowth but it might improve the crown. If 5 cm works still best, stick to this.
  • low quality = check this one. It is unchecked standrad

Ignore the last one, I have to remove from the dialog. The segmentation might be the next step I try to publish as manuscript and I expect still some work from my side on this step during the writing process.

You might have one of the most extreme plots to test my program. I am not sure how well the pipeline works here.


RE: Paramétrage segmentation/ décimation - Added by Domenech Anthony about 3 years ago

Good evening Mr. Hackenberg Jan,

Thank you for your answers.

1) I tested your recommendations ...
It turns out that in our case the cloud decimated in 5cm is segmented better than the 2cm.

This seems incoherent but the result proves it.
Here are the parameters: seep computing 01 and rank 0.2 and low quality.

If you have an explanation I would like to know.
For I am not of explanation.

2) I would like to have your opinion on using the results of a segmentation (from decimation 5cm) to carry out a second segmentation with a cloud of point decimate 2cm?

1st segmentation: cluster + decimated vegetation 5cm
2nd segmentation: result 1st + vegetation or decimated vegetation 2cm.

You will understand, my drums are segmented thanks to the 5cm decimation now I am trying to restore height density so that our algorithm can segment properly.

Thank you for your availability.

RE: Paramétrage segmentation/ décimation - Added by Hackenberg Jan about 3 years ago

Hello Anthony
I am not sure if I understand everyhing correctly.

First if not done yet, test if you can improve by stting the speed to 5 (test 3 if 5 is too slow first) and keeping range with 0.2 and low quality checked.

5 cm might be the point to point distance for the crown points. You have expectedly high trees and the resolution drops of course with increasing distance. If you scanned with lower resolution, than thats it. 5cm averages better for the complete scene than 2 in that case and my algorithms can run smoother in that case.

If I understood you want to transfer the segmentation of 5cm to the 2cm cloud. I can provide a step for this, np. But I will be on vacations after this post for one week.