SemCityLoc uses semantic 3D city models to boost aerial localization accuracy 4x

Published as an ECCV 2026 paper, SemCityLoc (by Jingfeng Mao and 7 co-authors) tackles a critical bottleneck in aerial navigation: accurate 6DoF localization without relying on expensive radiometric 3D reconstructions or high-quality GNSS signals. Instead, the system uses foundation-model-derived semantic features and monocular depth maps to align with lightweight, LoD-compliant (Level of Detail 1-3) semantic 3D building models. The key insight is that semantic surface geometry—such as walls, roofs, and ground planes—provides more discriminative and scalable constraints than traditional contour or texture matching, especially in repetitive or occluded urban environments.

To rigorously evaluate the method, the authors introduce SemCityLockeD, the first real-world benchmark combining centimeter-accurate UAV pose ground truth with standardized LoD1 to LoD3 city models and low-altitude imagery. Experiments show that SemCityLoc significantly outperforms existing map-based localization approaches: recall improves by up to 36%, and mean positional error drops from 9.89m to just 2.62m in dense urban canyons. The framework is fully open-source, with code and data publicly available, and its lightweight semantic priors suggest a path toward real-time onboard deployment for autonomous drones in GPS-denied environments.