LaMP: Last-Meter Manifold Pose Estimation for Image-Goal Terminal UAV Navigation
Abstract
Image-goal terminal UAV navigation guides a UAV from its current viewpoint to a goal viewpoint specified by a reference image. Local feature matching is brittle under the large viewpoint change between the two views, and monocular geometry alone struggles to supply the target-centric metric range the task demands. We propose LaMP (Last-meter Manifold Pose estimation), a framework that builds on the manifold prior induced by the target-centric terminal approach and injects it through supervision rather than as a loss constraint. Specifically, it lifts the task from predicting its two scalar outputs, heading and range change, to estimating the full 6-DoF relative pose, with labels constructed in closed form from the trajectory geometry. A geometry-supervised pose head reads heading from its continuous rotation, and a lightweight range head trained with a loss matched to the task’s relative-error metric supplements it for metric range. Both heads share a frozen VGGT backbone that encodes the two views jointly. LaMP attains sub-degree heading and sub-meter range error on held-out buildings and an official error of 0.0024 on the PairUAV hidden test (5th on the final leaderboard). Across unsampled viewpoints, altered approach slopes, and real drone imagery, heading remains accurate while metric range follows the assumed approach geometry. Code is publicly available at https://github.com/Polaris-F/vggt-pairuav.