A Survey on Visual Geolocalization: Taxonomy, Progress, and Prospects

Source abstract

Visual geolocalization aims to estimate the geographic location where an image was captured solely from its visual content. By enabling location inference without GPS signals or geographic metadata, it supports a wide range of applications, including autonomous navigation, urban computing, and location-based services. Over the past decade, visual geolocalization has witnessed remarkable progress driven by advances in visual representation learning, large-scale geo-tagged datasets, and, more recently, large vision-language models (LVLMs), giving rise to increasingly diverse methodological paradigms and problem settings. However, existing surveys primarily focus on specific application scenarios or traditional geolocalization approaches, leaving recent developments in reasoning-based geolocalization, multimodal foundation models, and emerging benchmarks insufficiently covered. In this survey, we provide a comprehensive review of visual geolocalization from the perspectives of both datasets and methodologies. We first present a unified taxonomy that organizes existing methods into three representative paradigms: retrieval-based, prediction-based, and reasoning-based localization. Building upon this taxonomy, we systematically review representative algorithms, benchmark datasets, evaluation protocols, and practical applications, while highlighting the evolution of the field and the relationships among different methodological paradigms. Furthermore, we discuss current challenges and identify promising future research directions, with particular emphasis on multimodal reasoning, agentic geolocalization, and open-world localization. We hope this survey provides a unified understanding of the rapidly evolving field and serves as a valuable resource for both newcomers and experienced researchers.