ARIA-ReID: Altitude-Robust Identity Association for Aerial-to-Ground Person Re-Identification

Kaustubh S. Bukkapatnam, Laksh Patel, Siddharth Karuturi

Track: Non-Proceedings Track

Keywords: aerial person re-identification, altitude-conditioned normalization, contrastive learning

TL;DR: Altitude-conditioned features and contrastive loss improve aerial-to-ground person re-ID robustness and accuracy on AG-ReID.

Abstract: Matching persons observed by a UAV at altitude against a ground-level gallery poses a fundamentally harder domain gap than conventional person re-identification: viewpoint varies from near-nadir to oblique, apparent resolution drops with altitude, and atmospheric turbulence blurs fine discriminative detail. Existing re-identification methods—designed for near-horizontal cross-camera matching—degrade sharply above 30 m altitude. We present ARIA-ReID (Altitude-Robust Identity Association for Aerial-Ground Re-Identification), a framework with two complementary components: (1) an Altitude-Conditioned Normalizer (ACN) that learns feature re-weighting as an explicit function of estimated altitude and viewing angle, and (2) a Cross-View Contrastive (CVC) training objective with a provably tighter alignment bound than standard InfoNCE when the query and key originate from different viewpoint distributions. On the AG-ReID benchmark , ARIA-ReID achieves Rank-1 = 78.6% and mAP = 67.4%, outperforming the strongest baseline by +12.8% mAP. Performance degrades gracefully with altitude (+9.7 pp Rank-1 advantage at 120 m vs. the strongest competitor), confirming that ACN provides the altitude-specific invariance that prior methods lack.

Submission Number: 11