Efficient knn on large points sets

[QingyongHu]

Thanks for your code.

It seems that knn_bruteforce function only support point sets less than 8k. As the warning shows:

"point sets greater than 8k are not yet supported!! Change to knn_graph operation!! "

Is there any efficient way to perform knn on large point set? (e.g. when the number of points is 100K or even more), otherwise, I still have to split the scene as many blocks or sample a part of points from the whole scene each time. However, this will undoubtedly lose the geometry information.

[PatWie]

This is ongoing work we haven't published yet. Note, in the paper we had a static scene such that a simple kd-tree beforehand did the job.

But, we already have a new network-layer that can compute these neighborhoods on-the-fly for larger point-clouds. Stay tuned :-)

[MacOS]

Thank you for your work and publishing your code!

Note, in the paper we had a static scene such that a simple kd-tree beforehand did the job.

I'm curious, which implementation of kd-tree did you use?

But, we already have a new network-layer that can compute these neighborhoods on-the-fly for larger point-clouds. Stay tuned :-)

Can you give me any time frame within you plan to make this public?

Thank you in advance!