Tracing locations from pixels: A survey of worldwide image geo-localization

Worldwide image geo-localization (WIGL) estimates the geographic coordinates of a photograph from its visual content, without metadata on the query image and without a known search area. This capability supports practical applications, including cybercrime source tracing, disaster response, and autonomous navigation. In this manuscript, a comprehensive review of WIGL is presented, in which current methods are systematically organized into four paradigms: classification-based, retrieval-based, generation-based, and retrieval-augmented generation (RAG). For each paradigm, representative approaches and core design principles are examined, with their strengths and limitations compared. Training, testing, and fine-tuning datasets are also reviewed in depth. A spatial distribution analysis of these datasets reveals notable coverage biases in existing benchmarks. Standard evaluation metrics are consolidated, and the reported results on six test sets are aggregated, providing a common basis for comparison across methods. The comparative analysis of reported results indicates that RAG approaches currently achieve the highest accuracy on major benchmarks, though this advantage is partly confounded by larger backbone models and training data while generation-based approaches offer a promising alternative for zero-shot location inference. Finally, several under-explored research challenges that are rarely discussed in prior surveys are identified and outlined. These include the difficulty of disambiguating visually similar images captured at distant locations, as well as other open problems to guide future work.