Learning Global Coordinate Regression with Prior-Guided Fusion for Large-Scale LiDAR Relocalization
Lists
Only you see this
Abstract
Accurate and robust LiDAR relocalization in large-scale outdoor environments requires a method capable of inferring globally consistent 6-DoF poses even under challenging conditions such as dynamic objects, partial occlusions, and scene appearance changes. While retrieval-based pipelines have made notable progress, regression-based approaches that directly map point clouds to global coordinates offer superior efficiency and end-to-end trainability. In this work, we propose a temporal prior-based localization framework that combines point-wise global coordinate regression with a learned prior generation mechanism to improve pose accuracy and scene-level consistency. At its core, a global coordinate estimation module estimates per-point 3D global positions alongside uncertainty weights within a learned map representation. A prior coordinate generation module leverages historical scan information to generate inter-frame point-level priors via transformer-based attention, capturing the geometric continuity that single-frame regressors inherently ignore. An uncertainty-guided coordinate fusion module then merges current predictions with generated priors, weighting contributions by predicted uncertainty to suppress noise from dynamic or occluded regions. The fused coordinates are passed to a robust pose solver to yield the final 6-DoF localization result. Extensive experiments on the NCLT and Oxford RobotCar benchmarks confirm that our method achieves state-of-the-art relocalization accuracy, with particular gains in long-term dynamic scenes where single-frame regression methods degrade significantly.
BibTeX
@misc{zhao2026learning,
author = {Xibin Zhao},
title = {{Learning Global Coordinate Regression with Prior-Guided Fusion for Large-Scale LiDAR Relocalization}},
year = {2026},
url = {https://openreview.net/forum?id=TEcFRR9hyc},
howpublished = {OpenReview},
}Loading related content...