Cross-View Geo-Localization

Cross-view geo-localization uses images taken from different viewpoints—such as a street-level photo and a satellite image—to estimate location and sometimes orientation, supporting navigation when GPS is unavailable. Common approaches learn image embeddings for contrastive retrieval or hierarchical map search, while geometric methods use 3D structure, bird’s-eye-view features, landmarks, or depth to align views; vision-language models can also help rerank candidates. Systems increasingly combine multiple frames with inertial sensors or map information to improve accuracy and reliability. A central challenge is generalizing across changes in viewpoint, altitude, weather, and sensor type while balancing precise localization against data, computation, and annotation costs.