search

flann-lib / flann

Fast Library for Approximate Nearest Neighbors
http://people.cs.ubc.ca/~mariusm/flann
Other
2.25k stars 648 forks source link

CUDA strikes #330

Closed wbrandenburger closed 5 years ago

wbrandenburger commented 7 years ago

Hi,

i have some troubles with the CUDA implementation of FLANN. I think the problems have something to do with the memory. And there are another problems which indicate that there is still a lot of room for improvements Im going to reimplement the CUDA-implementation. In this connection I will use as much as possible from your library. Are you still interested in my solution?

Greets from Munich

mariusmuja commented 7 years ago

Hi,

Yes, of course, a contribution that improves the CUDA implementation is welcome.

wbrandenburger commented 7 years ago

Hey Marius

thanks for the quick response! I think I have understood it now better. You have to know that Im a geodesist and I prefer more accuracy than fast and less exact solutions :) Moreover I work as far as possible in low dimensions. So the kdtreesingleindex offers the best solution for me. In the searchalgorithm of this index the distance from the querypoint to the respective node will be computed in the expected way and produces therefore the correct solution. But I think the way to compute mindist and new_distsq in kdtreeindex is debatable. I think I will test it so that I will return in peace with myself.

I want to take this opportunity to commend the clarity of this library. It is so straightforward to work through the library. Thanks a lot! It would be a enormous pleasure for me to be a part of FLANN.

I am working at the University of the Federal Armed Forces in Munich, Germany. At the institute of applied computer science I have the task to implement a efficient moving least squares algorithm to reduce hugh pointclouds. I will try to use a processingpipeline on a graphiccard to see whether it has benefits or not. So I need a slim, fast and accurate kdtreesearch in CUDA. If I can share my solution with the community it will be a make me a little bit happier :)

Im looking forward to hearing from you...

Am 13.03.2017 um 03:14 schrieb Marius Muja:

Hi,

Yes, of course, a contribution that improves the CUDA implementation is welcome.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/mariusmuja/flann/issues/330#issuecomment-285999861, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AYmPXLXnuiCj8y4-XbarzeVrm8EGP8Oaks5rlKZ0gaJpZM4Mai2G.

-- Wolfgang Brandenburger

Mainstraße 15f 85579 Neubiberg