Open-Vocabulary Object Detection in UAV Aerial Remote Sensing Images Based on YOLOE

Abstract

Object detection in unmanned aerial vehicle (UAV) aerial remote sensing images is an important application direction of computer vision, widely used in urban traffic monitoring, emergency rescue, agricultural plant protection, and other fields. Unlike traditional ground-level perspectives, UAV aerial images present challenges including large target scale variations, dense small objects, and complex backgrounds. Moreover, traditional closed-set detectors cannot identify unknown categories unseen during training. To address these issues, this paper performs a two-stage training adaptation of the off-the-shelf YOLOE (You Only Look Once - Everything) architecture on the VisDrone2019 dataset. RepRTA re-parameterization, SAVPE visual prompt encoding, and lazy region-prompt contrastive learning are core modules proposed in the original YOLOE paper, which are reviewed here only for a self-contained presentation and are not claimed as contributions of this work. Experimental results demonstrate that YOLOE-11-L achieves 48.6 % mean Average Precision (mAP@0.5) on the VisDrone2019 test-dev set, with an inference speed of 161 Frames Per Second (FPS), while supporting zero-shot detection capabilities via both text prompts and visual prompts. Notably, significant accuracy improvements are observed for small-object categories such as pedestrians and bicycles, while simultaneously satisfying real-time performance and open-vocabulary recognition requirements, providing an effective technical solution for UAV remote sensing inspection.