Expansion to point-to-voxel correspondence?

[dch4o]

Hi all,

First of all, thank you for sharing your work and for all the effort you've put into maintaining such a powerful and accessible framework.
The design and core ideas behind kiss-icp inspired me to revisit the fundamentals of the ICP algorithms.

I have been applying kiss-icp in various environments, both indoor and outdoor,
and during this process, I started to wonder a possible expansion:
replacing the point-to-point distance calculation with a point-to-voxel distance metric, similar to gicp.

While I understand that the original paper and design of kiss-icp are participating in point-to-point ICP, I believe it is a bit of a shame not to explore point-to-voxel correspondence, especially given the use of a voxel hash map in the local map representation.
I believe this extension can enhance odometry estimation by improving correspondence stability, especially in indoor environments where walls, floors, and other rectangular structures are common.

If this idea aligns with your interests, I would be happy to make a draft implementation in a few days.
If not, would it be possible for me to create a new repository (perhaps named kiss-gicp) based on your current work?

Looking forward to your thoughts.
Thanks in advance!

[benemer]

Do you mean maintaining a per-voxel distribution for the local map points similar to the normal distributions transform? Or do you mean using the distance of a point to a voxel center as an error function?

The first one is certainly something one could exploit. However, we aimed for a more generalized performance at maximum speed, and adding and maintaining per-voxel distributions will probably increase the complexity and runtime.

Feel free to give it a try! I think it is better to fork this repo instead of opening a PR here, because this is a more conceptual change that requires a lot of benchmarking before we could consider replacing our point-to-point ICP :)

Let us know how it went!