Marrying Feed-Forward 3D Reconstruction with Fine-Grained UAV Pose Alignment
Abstract
Last-meter UAV navigation requires predicting the relative pose between a source view and a target view under severe scale ambiguity and limited co-visibility, a regime in which cross-view retrieval and appearance-based matching pipelines struggle. By treating this fine-grained pose alignment as an inherent two-view geometry problem, we propose a simple but effective framework that leverages a pretrained 3D feed-forward reconstruction model as the backbone for pairwise UAV pose estimation. Our framework attaches two lightweight regression heads to directly predict the relative heading and range required to align the drone with the target pose. Despite its simplicity, our method ranks 4th in the UAVM’26 PairUAV challenge and significantly surpasses the target-driven navigation baseline on both heading and range metrics, demonstrating that feed-forward 3D reconstruction models provide effective geometric priors for fine-grained UAV pose alignment in challenging scenarios. Our code is available at https://github.com/Xyzzz1/UAVM26-Solution.