L2P-Former: Rehearsal-Free Prompt Routing for Continual LiDAR Place Recognition

Abstract

LiDAR place recognition (LPR) supports long-term localization and loop closure, yet models deployed in changing environments must learn new places without forgetting previous ones. Fine-tuning adapts effectively but degrades performance on earlier domains, while existing continual LPR methods typically require memory replay and knowledge distillation. We present L2P-Former, a rehearsal- and distillation-free prompt-routing framework built on a frozen MinkLoc3D encoder. A scene query retrieves key–value prompts from a learnable pool, and a frequency-aware criterion combines feature similarity with historical prompt usage to promote exploration during incremental training. The retrieved prompts are refined by self-attention and guide sparse local features through Point-Query/Prompt-KV cross-attention before frozen generalized-mean (GeM) aggregation. Only the prompt module is optimized. At Step 1, all methods use the shared Oxford-pretrained base model; from Step 2 onward, every observed domain, including Oxford, follows the same prompt branch without domain labels. On the four-stage continual benchmark, L2P-Former achieves a final average recall at rank one of 77.22% and a forgetting score of 7.59%. Relative to FineTune, it gains 4.73 points in final accuracy and reduces forgetting by 14.45 points, achieving a favorable stability–plasticity balance without storing historical point clouds.