MICL: Mutual Information Guided Continual Learning for LiDAR Place Recognition

Abstract

LiDAR Place Recognition (LPR) aims to identify previously visited places across different environments and times. Thanks to the recent advances in Deep Neural Networks (DNNs), LPR has experienced rapid development. However, DNN-based LPR methods may suffer from Catastrophic Forgetting (CF), where they tend to forget previously learned domains and focus more on adapting to a new domain. In this letter, we propose Mutual Information-guided Continual Learning (MICL) to tackle this problem in LPR. We design a domain-sharing loss function Mutual Information Loss (MIL) to encourage existing DNN-based LPR methods to learn and preserve knowledge that may not be useful for the current domain but potentially beneficial for other domains. MIL overcomes CF from an information-theoretic perspective including two aspects:1) maximizing the preservation of information from input data in descriptors, and 2) maximizing the preservation of information in descriptors when training across different domains. Additionally, we design a simple yet effective memory sampling strategy to further alleviate CF in LPR. Furthermore, we adopt adaptive loss weighting, which reduces the need for hyperparameters and enables models to make optimal trade-offs automatically. We conducted experiments on three large-scale LiDAR datasets including Oxford, MulRan, and PNV. The experimental results demonstrate that our MICL outperforms state-of-the-art continual learning approaches.