KFGOD: A Fine-Grained Object Detection Dataset in KOMPSAT Satellite Imagery
2025
Dong Ho Lee | Ji Hun Hong | Hyun Woo Seo | Han Oh
Object detection in high-resolution satellite imagery is a critical technology for various applications, yet it faces persistent challenges due to extreme variations in object scale, orientation, and density. The development of numerous public datasets has been pivotal for advancing the field. To continue this progress and expand the diversity of sensor data available for research, we introduce the KOMPSAT Fine-Grained Object Detection (KFGOD) dataset, a new large-scale benchmark for fine-grained object detection. KFGOD is uniquely constructed using 70 cm and 55 cm resolution optical imagery from the KOMPSAT-3 and 3A satellites, sources not covered by existing major datasets. It provides approximately 880,000 object instances across 33 fine-grained classes, encompassing a wide range of ships, aircraft, vehicles, and infrastructure. The dataset ensures high quality and sensor consistency, covering diverse geographical regions worldwide to promote model generalization. For precise localization, all objects are annotated with both oriented (OBB) and horizontal (HBB) bounding boxes. Comprehensive experiments with state-of-the-art detection models provide benchmark results and highlight the challenging nature of the dataset, particularly in distinguishing between visually similar fine-grained classes. The KFGOD dataset is publicly available and aims to foster further research in fine-grained object detection and analysis of high-resolution satellite imagery.
Показать больше [+] Меньше [-]Ключевые слова АГРОВОК
Библиографическая информация
Эту запись предоставил Multidisciplinary Digital Publishing Institute