Foundation and Frontiers in Visual Place Recognition: Advancing Architecture Evaluation and Adaptive Memory Learning

Visual Place Recognition (VPR) refers to the task of identifying the geographic or semantic location depicted in a generic image. This problem is a particularly crucial component for Simultaneous Localization and Mapping (SLAM) in autonomous robotics in an environment in which GPS is not available, as it allows the loop closure algorithm for path correction. However, the practical use of VPR systems is constrained by conventional metrics used for model evaluation, generalization across changing conditions, and the impossibility of complete model retraining on resource-constrained devices. This work deals with the limitations of standard ranking metrics in VPR by introducing an evaluation methodology based on the analysis of cosine similarity distributions, focusing on Average Precision (AP) to take into account the model's discriminative ability; finally, it proposes and validates a model-agnostic improvement for a continual learning strategy, where the proposed intelligent memory management system improves performance with minimal retraining epochs.