EAGLE: Large-Scale Vehicle Detection Dataset in Real-World Scenarios using Aerial Imagery

Azimi, Seyed Majid; Bahmanyar, Reza; Henry, Corentin; Kurz, Franz · 2021 · Crossref

DOI: 10.1109/icpr48806.2021.9412353

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper introduces EAGLE (oriEnted vehicle detection using Aerial imaGery in real-worLd scEnarios), a large-scale dataset designed to address the scarcity of diverse, high-quality data for multi-class vehicle detection in airborne imagery. While object detection in ground imagery has advanced significantly, progress in aerial domains remains limited due to the lack of datasets that reflect real-world complexities such as varying weather, illumination, occlusion, and object orientation. The authors aim to provide a robust benchmark for traffic monitoring, disaster management, and urban planning applications, where accurate vehicle localization and orientation estimation are critical. The EAGLE dataset comprises 8,820 high-resolution aerial images (936 × 936 px patches from larger stitched images) acquired between 2006 and 2019 across five countries. Images were captured from altitudes of 300–3000 meters using DSLR cameras under diverse conditions, including snow, haze, rain, and varying times of day. The dataset contains 215,986 annotated vehicle instances, categorized into small vehicles (e.g., cars, vans) and large vehicles (e.g., trucks, buses). Annotations utilize oriented bounding boxes (OBBs) defined by four corner coordinates and an orientation angle, along with difficulty flags for visibility and orientation clarity. To evaluate performance, the authors define three detection tasks: horizontal bounding boxes (HBB), rotated bounding boxes (RBB), and OBB. They establish two dataset splits: a random split and a more challenging campaign-based split that ensures no overlap in flight missions between training and testing sets. Experiments benchmarked state-of-the-art detection algorithms, including Cascade Mask-RCNN, Faster R-CNN, and YOLOv3. Results indicate that EAGLE presents significant challenges for existing models. Cascade Mask-RCNN achieved the highest performance, with 39.29% mean Average Precision (mAP) for HBB detection. Adapting the model for RBB detection improved performance to 37.23% mAP, while the proposed OBB approach achieved 39.29% mAP with 67.34% angle accuracy. The study demonstrates that OBBs provide superior localization precision compared to HBBs and RBBs, particularly in densely packed scenarios where horizontal boxes cause excessive overlap. Furthermore, models trained on EAGLE showed reduced performance when tested on the DOTA dataset, highlighting domain-specific challenges. The significance of this work lies in providing the largest and most comprehensive dataset for aerial vehicle detection to date, enabling the development of practical algorithms for real-world scenarios. By including diverse environmental conditions and precise orientation annotations, EAGLE supports research not only in detection but also in related tasks such as haze removal, shadow handling, and super-resolution. The established baselines and challenging splits offer a rigorous framework for future advancements in airborne object detection.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms