Investigate Adaptive Threshold

[tizianoGuadagnino]

Motivation

Our current implementation, based on the discussion in #15 is more consistent with using wheel odometry as an initial guess. However, many different implementations will work in this scenario; for example, @benemer proposes inferring from the map parameters as simple.

$\tau = 6 \cdot \frac{\text{voxel size}}{\sqrt{\text{max points in voxel}}}$

According to our evaluation, this gives identical results. This raises the question of how we can properly decide on a threshold value just based on data. Our main issue is a lack of proper data to test this, and we will need lots of data to make a final decision based on a bunch of different scenarios. @benemer and I want to come back to this, and anyone is welcome to this discussion in this issue or future PRs.