Towards Robust DEM-Based Monocular Depth Rescaling for UAVs: A Systematic Analysis

Monocular depth estimation suffers from inherent scale ambiguity, limiting its real-world applicability. While recent foundation models produce high-quality relative depth maps, their predictions lack metric consistency, which becomes critical in aerial environments. Recent rescaling approaches address this limitation by leveraging Digital Elevation Maps (DEM) as absolute references, yet existing DEM-based methods remain difficult to deploy due to their reliance on offline ground segmentation techniques and the absence of standardized validation frameworks. To overcome these challenges, we introduce a Synthetic-to-Geo-Real simulation framework that jointly models realistic flight dynamics and real georeferenced terrain, enabling a novel, systematic, and controlled evaluation of DEM-based rescaling across diverse operational conditions. The framework provides a unified benchmarking setting to compare multiple state-of-the-art depth models and alternative ground segmentation strategies under consistent assumptions. Our results show that semantic segmentation provides more consistent performance across models, altitudes, and environments, achieving up to a 60% reduction in error over geometric segmentation. We further observe that decoupling segmentation from depth estimation and selectively sampling ground points can improve scale recovery. These findings support the practical viability of DEM-based metric depth rescaling for Unmanned Aerial Navigation.