Clarification Needed on Weight Initialization Strategy for DRPM Integration in SLAM Systems

[BingyangZHOU]

Amazing works!
Thanks for the great work and code. I am trying to integrate drpm into my SLAM system. However, I am still quite confused for the weight setting after reading the paper and code(i.e., weights_squared defined in data.h).
When I get point and estimated normal from my SLAM system, how can I set the weight？Can I just set them all a const number like 1/(number of point)?
Any guidance or minimal code examples would greatly accelerate my integration efforts. Thank you for your time!

[johhat]

Thank you! Regarding weights, referring to the paper.

• $w_{r_i}$ are user-specified per-residual weights. E.g. for whitening the residual. If you want to give all points equal weight, this can be set to a const number (e.g. 1.0) for all residuals. If you want some residuals to be given more weight than others in the optimization, the weights should change accordingly.
• $w_{\rho_i}$ is the attenuation induced by the choice of (robust) cost function. The paper mentions the standard L2 norm and the Geman-McClure cost function.

A lot of choices are possible for residual weighting and (robust) cost function. The paper does not dictate any particular choice. The simplest place to start is to set $w_{r_i} = w_{\rho_i} = 1$, but in a real-world slam system one would typically use a robust cost function (resulting i non-unity $w_{\rho_i}$).