Cross-Modal Enhancement and Benchmark for UAV-based Open-Vocabulary Object Detection (2509.06011v1)

Abstract: Open-Vocabulary Object Detection (OVD) has emerged as a pivotal technology for applications involving Unmanned Aerial Vehicles (UAVs). However, the prevailing large-scale datasets for OVD pre-training are predominantly composed of ground-level, natural images. This creates a significant domain gap, causing models trained on them to exhibit a substantial drop in performance on UAV imagery. To address this limitation, we first propose a refined UAV-Label engine. Then we construct and introduce UAVDE-2M(contains over 2,000,000 instances and 1800 categories) and UAVCAP-15k(contains over 15,000 images). Furthermore, we propose a novel Cross-Attention Gated Enhancement Fusion (CAGE) module and integrate it into the YOLO-World-v2 architecture. Finally, extensive experiments on the VisDrone and SIMD datasets verify the effectiveness of our proposed method for applications in UAV-based imagery and remote sensing.