Segmentation of big braches from a single tree

[Zihui]

Hello everyone,

I am doing some reconstruction of single trees from TLS data. I would like to exclude the main trunk and the big branches from my point clouds(i.e. only leaved points remaining).


The picture is a small part of the point clouds of a single tree.

I have tried to filter out the branches by RGB information and cylinder shape detection, respectively. But neither of them worked well.

Is there any solution for this?

Thank you in advance.

Zhu

[daniel]

Maybe with Canupo?

When detecting cylinders, how bad the results were? (can you maybe post the image again? it seems to have failed the first time)

[Zihui]

Thank you Daniel for responding! :)


I hope those pictures are available now :)


Fig.1 Original point clouds of a single tree


Fig.2 Cylinder detection( many errors)


Fig.3 CANUPO classified


Overall, CANUPO has a better performance for excluding the big branches. However, I will lose many points of leaves, which is missing my goal.

Should I train the CANUPO classifier more carefully that I need to provide preciser sample classes?

Also, it was taking some time (20min to a few hours) to classify a 500MB data, occupying 90% CPU. (both for i7 7700,32GB and Xeon W2145, 128GB)


Best regards
Zhu

[daniel]

Yes, if you want to differentiate the trunk, big branches, small branches, and leaves, you'll have to be very careful during the classifier(s) design. You normally have to define several classifiers (one to classify the leaves, and everything else, then one to classify the small branches, and the remaining that you will apply to the points remaining after applying the first classifier, etc.).