Go to ICLR 2025 Conference homepageClass-Relational Label Smoothing for Lifelong Visual Place Recognition
Visual Place Recognition (VPR) is a task of estimating the location of a query image, predominantly executed through image retrieval using learned global descriptors from a reference database of geo-tagged images. While recent approaches have aimed to improve the scalability of VPR training by leveraging classification loss as a proxy task, this leads to a task gap between classification and retrieval - classification discretizes the feature space into distinct class regions, often overlooking visual differences between classes. This gap makes VPR systems particularly vulnerable to extreme visual changes such as lifelong variations. To remedy these problems, we propose a novel Class-Relational Label Smoothing (CRLS) that transforms one-hot labels into soft labels by considering visual information of inter-class relations. We further enhance this method by dynamically adjusting the influence of CRLS based on the stability of class weights, which is quantified by their magnitudes. Importantly, our findings suggest that the magnitude of class weights serves as an indicator of class stability, which is also supported by derivative analysis. We demonstrate that our method outperforms state-of-the-art methods on the most extensive 17 benchmarks, effectively bridging the task gap between classification and retrieval in visual place recognition. Codes and trained weights will be made publicly available.