A Lightweight Universal Reranking-based Optimizer for Robust Lidar Place Recognition

Abstract

Lidar place recognition (LPR) is a cornerstone of robust localization for autonomous systems. However, existing methods relying on fixed-length global descriptors are vulnerable to challenging conditions like rotational shifts and long-term environmental changes. While reranking-based optimizers have been introduced to refine initial localization candidates, they are fundamentally constrained by operating on small, fixed-size candidate pools. This design becomes a critical bottleneck when the correct match is initially ranked too low by the descriptor to be included in this limited pool, rendering the reranking process ineffective. To address this bottleneck, we propose LPR-Mate, a lightweight and universal reranking-based optimizer. LPR-Mate introduces two key innovations. First, a dynamic trigger mechanism proactively restructures the candidate pool by using rapid geometric evaluation to identify and promote high-quality candidates from a wider search space, thus preventing the premature exclusion of true matches. Second, a dedicated reranking network uses a gated fusion mechanism to adaptively weigh and combine multi-level information—including local features, global descriptors, and spatial compatibility scores—for precise verification. Extensive experiments demonstrate that LPR-Mate boosts the average Recall@1 of baseline methods by 39.56% under perturbations and verifies the ground-truth match with over 97% recall, effectively correcting their initial localization errors. As a universal plug-and-play module, LPR-Mate seamlessly enhances diverse LPR architectures without retraining, ensuring both efficiency and robustness.