GEORANK: A VISUAL FEATURE–BASED IMAGE RANKING FRAMEWORK FOR GEOLOCATION PREDICTION

Abstract

We propose GeoRank, a unified framework that extracts multi-level visual features from real-world images, ranks reference images by similarity, and predicts geographic coordinates (latitude, longitude) via a ranking-aware geolocation module. GeoRank fuses object-level cues (detected by a modern detector), global scene embeddings (CLIP / ViT), and landmark-sensitive descriptors (GeoCLIP-style alignment) into a single geolocation-sensitive embedding. A FAISS-backed nearest-neighbor retrieval and ranking stage produces top-k candidate locations, and the final prediction is obtained by a weighted regressor over the ranked candidates. In experiments on established geolocation benchmarks, GeoRank improves localization accuracy over a CLIP-only baseline and shows robustness to occlusions and scene variation. Key contributions: (1) an interpretable feature fusion and ranking pipeline for geolocation, (2) a reproducible implementation and Streamlit demo architecture, and (3) a thorough literature survey and evaluation plan.