VIPeR: Visual Incremental Place Recognition With Adaptive Mining and Continual Learning

Abstract

Visual place recognition (VPR) is essential to many autonomous systems. Existing VPR methods demonstrate attractive performance at the cost of limited generalizability. When deployed in unseen environments, these methods exhibit significant performance drops. Targeting this issue, we present VIPeR, a novel approach for visual incremental place recognition with the ability to adapt to new environments while retaining the performance of previous ones. We first introduce an adaptive mining strategy that balances the performance within a single environment and the generalizability across multiple environments. Then, to prevent catastrophic forgetting in continual learning, we design a novel multi-stage memory bank for explicit rehearsal. Additionally, we propose a probabilistic knowledge distillation to explicitly safeguard the previously learned knowledge. We evaluate our proposed VIPeR on three large-scale datasets—Oxford Robotcar, Nordland, and TartanAir. For comparison, we first set a baseline performance with naive finetuning. Then, several more recent continual learning methods are compared. Our VIPeR achieves better performance in almost all aspects with the biggest improvement of 13.85% in average performance.