Pose optimization for driving datasets

The evaluation is done with the poses from different pose optimization methods.
The evaluation metrics are from 3 categories to avoid any bias.
Novel View Synthesis: Nerfacto with LiDAR depth maps and Splatfacto with SfM+LiDAR initialization.
Structure-from-Motion: COLMAP reprojection error and track length.
Geometric consistency: Precision with the percentage of LiDAR points under 15cm of distance with the Delauney mesh, and average distance.
The results for each dataset are the average values over 5 selected sequences.

Pose optimization for autonomous driving datasets using neural rendering models