A Multi-Modal Per-Pair Pipeline for UAV Pose Estimation in Orbital Image Sets
Abstract
We present a per-pair UAV pose-estimation pipeline for orbital image sets built around a multi-modal fusion architecture: three frozen vision foundation models — DINOv2 visual semantics, SuperPoint+LightGlue correspondences, and DepthAnythingV2 depth statistics — are fused through a lightweight cross-attention head with a bounded, discontinuity-free pose parameterisation. The continuous estimates are then sharpened by a symmetry-aware four-stream test-time ensemble derived from the task’s exact pair-direction and mirror symmetries. Every hyperparameter is fixed by label-free criteria: ensemble weights follow from symmetry, and model selection uses training loss only. On the official UAVM 2026 PairUAV leaderboard (2.77M test pairs) our submission achieves a final score of 0.004023 (angular 0.00181, distance 0.006237); per-component leaderboard diagnostics show every stage of the pipeline contributing a measurable share.