More Clear, More Flexible, More Precise: A Comprehensive Oriented Object Detection benchmark for UAV

TL;DR

CODrone addresses a gap in UAV-oriented oriented object detection by providing a large-scale, high-resolution benchmark with 12 categories and oriented bounding boxes, captured across multiple cities, altitudes ($30$, $60$, $100$ m) and camera angles ($30^ 0^ 2 ilde ext{ and }90^ 0^ 2$ degrees). It benchmarks 22 methods, revealing robustness gaps in precise localization under varying viewpoints and distances, with $AP_{50}$ and $AP_{75}$ indicating performance dips at higher altitudes and oblique angles. The dataset comprises 10,004 images and 443,592 oriented annotations, with publicly available test annotations to enable reproducible evaluation, and highlights the need for rotation-aware and scale-adaptive approaches in real UAV deployments. Overall, CODrone provides a realistic platform to advance UAV perception, urban inspection, logistics, and disaster-response applications by promoting methods that generalize across altitude-angle variations and challenging urban scenes.

Abstract

Applications of unmanned aerial vehicle (UAV) in logistics, agricultural automation, urban management, and emergency response are highly dependent on oriented object detection (OOD) to enhance visual perception. Although existing datasets for OOD in UAV provide valuable resources, they are often designed for specific downstream tasks.Consequently, they exhibit limited generalization performance in real flight scenarios and fail to thoroughly demonstrate algorithm effectiveness in practical environments. To bridge this critical gap, we introduce CODrone, a comprehensive oriented object detection dataset for UAVs that accurately reflects real-world conditions. It also serves as a new benchmark designed to align with downstream task requirements, ensuring greater applicability and robustness in UAV-based OOD.Based on application requirements, we identify four key limitations in current UAV OOD datasets-low image resolution, limited object categories, single-view imaging, and restricted flight altitudes-and propose corresponding improvements to enhance their applicability and robustness.Furthermore, CODrone contains a broad spectrum of annotated images collected from multiple cities under various lighting conditions, enhancing the realism of the benchmark. To rigorously evaluate CODrone as a new benchmark and gain deeper insights into the novel challenges it presents, we conduct a series of experiments based on 22 classical or SOTA methods.Our evaluation not only assesses the effectiveness of CODrone in real-world scenarios but also highlights key bottlenecks and opportunities to advance OOD in UAV applications.Overall, CODrone fills the data gap in OOD from UAV perspective and provides a benchmark with enhanced generalization capability, better aligning with practical applications and future algorithm development.

Figures (7)

• Figure 1: Comparison of the proposed CODrone dataset with existing UAV oriented object detection (OOD) datasets. CODrone is designed as a UAV OOD benchmark for urban environments, constructed with high-resolution imagery, diverse object categories, and flexible variations in both altitude and camera angle. (The camera angle for UAV-ROD is not explicitly provided in the dataset and is therefore estimated based on available visual and metadata information.)
• Figure 2: Classification of existing remote sensing object detection benchmark based on image source and annotation method. The proposed CODrone addresses the scarcity of oriented object detection benchmarks for UAVs, while also introducing new challenges in terms of object categories, viewing angles, and other key aspects.
• Figure 3: Visualization of representative image examples. The scenes shown above, such as quays, highways, villages, snowy environments, and nighttime conditions, are rarely observed in existing OOD UAV datasets. CODrone includes, but is not limited to, these diverse environments, thereby offering stronger general ization capabilities.
• Figure 4: Visualization of annotated object categories in CODrone. We selected and labeled a wide range of object classes that are commonly encountered in urban environments and hold practical relevance for UAV applications. All instances are manually annotated with high-quality oriented bounding boxes.
• Figure 5: Visualization of object features at different resolutions. The left image shows a cropped region from an original CODrone image. The right side illustrates the appearance of the same objects under different downsampling scales. “1×” denotes the original resolution, while “2×” and “4×” correspond to 2× and 4× downsampling, respectively.