Learning from Mistakes: Loss-Aware Memory Enhanced Continual Learning for LiDAR Place Recognition

TL;DR

KDF+ addresses catastrophic forgetting in LiDAR place recognition by introducing loss-aware sampling and rehearsal enhancement within a continual learning framework. A loss predictor guides memory buffer construction by prioritizing harder samples, while a rehearsal loss progressively refines memory representations during new-task updates. The approach is validated across multiple LPR backbones and datasets, consistently outperforming baselines and even improving other CL-LPR methods when integrated. The results demonstrate robust retention of past knowledge with effective adaptation to new environments, offering a practical path for deploying CL-LPR in real-world autonomous systems.

Abstract

LiDAR place recognition plays a crucial role in SLAM, robot navigation, and autonomous driving. However, existing LiDAR place recognition methods often struggle to adapt to new environments without forgetting previously learned knowledge, a challenge widely known as catastrophic forgetting. To address this issue, we propose KDF+, a novel continual learning framework for LiDAR place recognition that extends the KDF paradigm with a loss-aware sampling strategy and a rehearsal enhancement mechanism. The proposed sampling strategy estimates the learning difficulty of each sample via its loss value and selects samples for replay according to their estimated difficulty. Harder samples, which tend to encode more discriminative information, are sampled with higher probability while maintaining distributional coverage across the dataset. In addition, the rehearsal enhancement mechanism encourages memory samples to be further refined during new-task training by slightly reducing their loss relative to previous tasks, thereby reinforcing long-term knowledge retention. Extensive experiments across multiple benchmarks demonstrate that KDF+ consistently outperforms existing continual learning methods and can be seamlessly integrated into state-of-the-art continual learning for LiDAR place recognition frameworks to yield significant and stable performance gains. The code will be available at https://github.com/repo/KDF-plus.

Figures (4)

• Figure 1: Comparison of different methods in mitigating catastrophic forgetting for LiDAR place recognition (LPR). At each step, the trained MinkLoc3Dkomorowski2021minkloc3d model is evaluated on the Oxfordmaddern20171 dataset, and both Recall@1 and Forgetting scores are reported. Solid lines represent Recall@1, while dashed lines indicate the Forgetting score. Higher Recall@1 values correspond to better place recognition performance, whereas lower Forgetting scores reflect stronger resistance to catastrophic forgetting.As shown in the figure, baseline methods such as InCloudknights2022irosincloud, CCLcui2023ralccl, and MICLliu2024micl experience substantial drops in Recall@1 and sharp increases in Forgetting as new tasks are learned. The previous method KDFwang2025ranking provides moderate improvements. In contrast, our proposed KDF+ consistently maintains the highest Recall@1 and achieves the lowest Forgetting score across all steps, demonstrating its superior capability in mitigating catastrophic forgetting while preserving recognition performance on previously learned tasks.
• Figure 2: Overview of the proposed KDF+ CL-LPR framework. The framework incorporates two key components: (1) Loss-Aware Sampling: A loss prediction layer estimates the difficulty or importance of samples from the previous dataset and produces a non-uniform sampling weight distribution (illustrated by the histogram). This enables prioritized selection of informative samples (e.g., LiDAR submaps) to populate the replay buffer, as indicated by the weighted samples. (2) Rehearsal Enhancement Loss: Applied during rehearsal when learning a new task, this loss encourages memory samples to further improve. As shown by the loss curves, it helps maintain performance on the previous task (yellow line) while the backbone adapts to the new task (green line), effectively reducing forgetting. In addition, the framework leverages ranking-aware distillationwang2025ranking to transfer knowledge from the previous backbone to the newly updated backbone.
• Figure 3: Detailed performance comparison under the 4-step protocol across three LPR backbones—PointNetVLADuy2018pointnetvlad, MinkLoc3Dkomorowski2021minkloc3d, and CASSPRxia2023casspr—and four continual learning methods: InCloudknights2022irosincloud, CCLcui2023ralccl, MICLliu2024micl, and KDFwang2025ranking. The results are presented as Recall@1 matrices, where each row corresponds to a continual learning method and each column corresponds to a different backbone. Within each matrix, Steps 1 through 4 report the Recall@1 performance on the environments encountered up to that step. Darker colors indicate higher Recall@1 values.
• Figure 4: Generalization of the proposed KDF+ components to other CL-LPR methods, InCloudknights2022irosincloud and MICLliu2024micl. Integrating the loss-aware sampling and rehearsal enhancement mechanisms into these baselines yields significant improvements in both mean Recall@1 and Forgetting Score.