Option for a start set instead of only a single start index?

[PhilippDahlinger]

Hello,

Thank you for your code; it works very well for me! I have a new project where I need to perform FPS on a point cloud, but I also need to ensure certain specific points are tracked. This requires that some points in the point cloud are always included in the sampled point cloud.

For a naive FPS, this can be achieved by including these specific points in the initial set, then iteratively calculating the farthest point from the current set and adding it to the set in each step.

Could this approach be implemented for more advanced FPS methods as well? The interface changes would be minimal; instead of specifying a single start index, it could allow for a list of start indices that are guaranteed to be included in the sampled point cloud. Ideally, these points would appear at the beginning of the output array for easy reference.

Would this feature be possible to add? I would love to use it!

Thanks again for your work!

[leonardodalinky]

Sure. The latest FPS algorithms still sample points iteratively and we could change the initial point sets. I would try to add this feature maybe in months. And I would be much appreciate if you could make certain PRs to change the interface and part of the internal logics.

[PhilippDahlinger]

That would be awesome. I can provide the Python interface update, however, I am not familiar with Rust so far, so I can't help you there much.

[PhilippDahlinger]

If you are more interested in implementing this feature in pytorch_fpsample, then I can also provide the interface logic there. I would obviously love to see this project done, so that I can use it in my work.

[h-ahl]

I would also be very interested in this feature. This library is excellent, but currently we need to work around the absence of this