EAGLE-ReID: Strategic Alignment and Delta Consistency for Extreme Far-Distance Aerial-Ground Re-Identification

Cheng-Jun Kang, Jin-Hui Jiang, Yu-Fan Lin, Chih-Chung Hsu; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 1618-1624

The proliferation of Unmanned Aerial Vehicles (UAVs) has revolutionized intelligent surveillance but introduced severe challenges for Person Re-Identification (ReID). The newly released DetReIDX benchmark highlights the "Extreme Far Distance" (XFD) problem, where targets are captured at variable altitudes (up to 120m) with drastic viewpoint shifts. In this paper, we present EAGLE-ReID, a robust framework securing the 3rd place in the VReID-XFD challenge. While adopting a Vision-Language Model (VLM) adapter architecture as our foundation, we diagnose a critical bottleneck via t-SNE analysis: the feature distributions between aerial and ground views are severely disjoint due to the extreme domain gap. To address this, we propose a strategic enhancement pipeline: (1) a Geometry-Aware Sampler (GAS) that forces the model to learn from sparse ground and abundant aerial pairs; and (2) a View-Manifold Delta (VMD) strategy to explicitly model the continuous transformation between views rather than enforcing rigid alignment. Our approach achieves an overall mAP of 29.00% on the official test set, demonstrating significant improvements in cross-view retrieval stability.