Implementation of FPFH and RANSAC in the code

[lemonci]

Hello again!

In the paper, you mentioned that

We use a coarse-to-fine registration approach to estimate loop edge constraints. For the coarse alignment, we apply the global registration method of Rusu et al. [31], which extracts Fast Point Feature Histograms (FPFH) from downsampled versions of the source ($P_s$) and target ($P_t$) point clouds. Correspondence search is then performed in the FPFH feature space rather than in Euclidean space. Optimization is embedded in a RANSAC framework to reject outlier correspondences, producing a rigid transformation of the source point cloud $S_s$ to align with the target $S_t$. Finally, ICP [3] is applied to the full-resolution point clouds to refine the coarse alignment estimate.

I do find the ICP is called here:
https://github.com/VladimirYugay/MAGiC-SLAM/blob/4ac5aa92ab70139546b42927b62e4f2fd6940222/src/utils/magic_slam_utils.py#L168C1-L172

However, the registration.init_transformation was just calculated by tracking instead of FPFH and RANSAC:
https://github.com/VladimirYugay/MAGiC-SLAM/blob/4ac5aa92ab70139546b42927b62e4f2fd6940222/src/entities/loop_detection/loop_detector.py#L161C1-L162

Could you explain a bit where FPFH and RANSAC were used?