Accuracy potential of visual localization exploiting high-end street-level imagery

Visual localization—the process of determining a camera's 6-degree-of-freedom (DoF) pose within a known scene—is a critical technology for autonomous navigation, robotics, and augmented reality. While GNSS is the standard for outdoor positioning, it is often unreliable in urban canyons due to signal occlusions and multipath effects. Although visual localization is a promising alternative, its potential to meet the sub-centimeter accuracy requirements of professional surveying has remained largely unverified, primarily due to the lack of large-scale, high-precision datasets with reliable ground-truth poses.

The authors introduce a scalable visual localization pipeline that uses georeferenced, high-resolution street-level imagery as the primary scene representation. Instead of relying on static 3D point clouds, the system performs on-the-fly local Structure-from-Motion (SfM) reconstruction using retrieved reference candidates, followed by PnP-based pose estimation. To validate this, they present the FHNW Muttenz dataset, which covers a 10 km street network mapped at two different time points. This dataset provides sub-centimeter ground-truth poses, allowing for a rigorous evaluation of localization accuracy under real-world conditions.

The evaluation demonstrates that visual localization can achieve median translation accuracies of 1–5 cm and rotation accuracies of 0.05–0.1°. Under optimal conditions, the system reached accuracies as low as 1 cm and 0.03°. These results indicate that camera-based localization is sufficiently accurate to complement or even replace survey-grade GNSS in challenging environments, enabling automated 3D geospatial data acquisition using standard consumer devices.

This research bridges the gap between academic visual localization methods and the high-precision demands of the surveying industry. By demonstrating that consumer-grade cameras can achieve survey-grade accuracy, the study paves the way for more accessible, automated, and cost-effective georeferencing and mapping solutions that do not depend on GNSS availability.