Understanding of "relocalization"-ish ability

[Lucasmogsan]

Hi,

First of all thanks for your interesting work and alternative approach to 3DGS SLAM using RGBD.

I'm currently benchmarking GS-ICP to other NERF or 3DGS based SLAM methods. One thing i notice is even when loosing track or drifting GS-ICP is able to "relocate" or adjust its localisation estimation... supposedly when revisiting something previously seen.

I'm quite curious on what causes this "relocalization" since the pipeline differs quite a bit from traditional SLAM implementations. Unlike loop closure and/or global as well as local bundle adjustment no optimization of previous poses are done as I understand. I guess this relocalisation-ish is then solely based on the GICP between the Target Gaussians (subset from the map) and the Soruce Gaussians (from current rgbd-frames point cloud).

Is this assumption correct, and could someone (@Riboha @bogus2000) explain this a little?