LPR-Mate: A Lightweight Universal Reranking-based Optimizer for LiDAR Place Recognition in Challenging Environments

Abstract. LiDAR place recognition (LPR) plays a critical role in simultaneous localization and mapping (SLAM) and autonomous driving systems. However, current LPR methods exhibit significant performance degradation under rotational shifts, noise interference, point cloud sparsity, and long-term environmental changes. This limitation stems from their reliance on fixed-length global descriptors, which lack the capacity to preserve comprehensive scene information in complex scenarios. To address these challenges, we propose LPR-Mate, a lightweight universal reranking-based optimizer that enhances the robustness of existing LPR frameworks in challenging environments. LPR-Mate processes top-k retrieval candidates from baseline LPR methods through a dual-stage pipeline: (1) A fast trigger mechanism evaluates spatial consistency between query and candidate scenes, selectively activating reranking only for low-confidence matches; (2) An independent reranking network refines candidate rankings by fusing local features, global descriptors, and spatial consistency scores through group and channel attention mechanisms. Extensive experiments on the Oxford RobotCar, NUS-Inhouse, and MulRan datasets demonstrate that LPR-Mate achieves >96% recall in localization accuracy validation and delivers a 32.34% average improvement in Recall@1 under rotational shifts, sparsity, and noise perturbations, while maintaining robustness for raw point clouds and long-term scenarios. As a plug-and-play module, LPR-Mate integrates seamlessly with diverse LPR architectures—including region-sampling and sparse-voxelization-based methods—without requiring retraining or structural modifications, ensuring computational efficiency and cross-architectural universality.