Abstract
Applications of unmanned aerial vehicle (UAV) in logistics, agriculturalautomation, urban management, and emergency response are highly dependent onoriented object detection (OOD) to enhance visual perception. Although existingdatasets for OOD in UAV provide valuable resources, they are often designed forspecific downstream tasks.Consequently, they exhibit limited generalizationperformance in real flight scenarios and fail to thoroughly demonstratealgorithm effectiveness in practical environments. To bridge this critical gap,we introduce CODrone, a comprehensive oriented object detection dataset forUAVs that accurately reflects real-world conditions. It also serves as a newbenchmark designed to align with downstream task requirements, ensuring greaterapplicability and robustness in UAV-based OOD.Based on applicationrequirements, we identify four key limitations in current UAV OOD datasets-lowimage resolution, limited object categories, single-view imaging, andrestricted flight altitudes-and propose corresponding improvements to enhancetheir applicability and robustness.Furthermore, CODrone contains a broadspectrum of annotated images collected from multiple cities under variouslighting conditions, enhancing the realism of the benchmark. To rigorouslyevaluate CODrone as a new benchmark and gain deeper insights into the novelchallenges it presents, we conduct a series of experiments based on 22classical or SOTA methods.Our evaluation not only assesses the effectiveness ofCODrone in real-world scenarios but also highlights key bottlenecks andopportunities to advance OOD in UAV applications.Overall, CODrone fills thedata gap in OOD from UAV perspective and provides a benchmark with enhancedgeneralization capability, better aligning with practical applications andfuture algorithm development.