Accuracy Potential of Visual Localization Exploiting High-End Street-Level Imagery

Researchers built a visual localization pipeline using precisely georeferenced street-level imagery and a new 10 km dataset with sub-centimeter ground truth.

Across multiple camera setups and scenes, median errors reached 1-5 cm in translation and 0.05-0.1 degrees in rotation.

Best-case performance was around 1 cm and 0.03 degrees, showing centimeter-level accuracy is achievable with high-end street imagery.

The results suggest visual localization can complement survey-grade GNSS for surveying, robotics, and geospatial data collection.