2026-10-12T09:20:57Z https://www.scidb.cn/oai

10.57760/sciencedb.j00240.00224 2026-06-14T14:31:11Z

2026-06-14 DIV: A visible-infrared object detection dataset for weakly aligned drone scenarios doi:10.57760/sciencedb.j00240.00224 en To support research on infrared-visible multimodal object detection from a drone perspective, a drone-based multimodal object detection dataset (DIV) has been constructed. Collected using real drone platforms, this dataset covers diverse scenarios, lighting conditions, and flight altitudes. It preserves cross-modal weak alignment characteristics in its design, providing a crucial experimental foundation for studying robust cross-modal fusion and object detection. This dataset comprises a total of 10926 images (5463 image pairs) and contains annotations for 87140 object instances. Annotation was performed using the roLabelImg tool to label objects in the images, covering five classes: car, van, truck, person, and non-motorized vehicle. The dataset covers both urban areas with high object density and mountainous and rural areas with open views and sparse objects, effectively simulating the requirements for UAVs to perform inspection and surveillance tasks in various environments. Based on the time of acquisition, we classify the data into three lighting conditions: daytime (60.64%), nighttime (28.67%), and low-light environments (10.69%). Night and low-light samples together account for nearly 40% of the total dataset, facilitating a comprehensive evaluation of the performance differences between visible and infrared modalities under complex lighting conditions. Regarding object scale, tiny objects smaller than 20 pixels constitute nearly 40% of the dataset, while small objects ranging from 20 to 32 pixels account for over 30%, providing ample samples for addressing the challenge of detecting small targets at long distances. Furthermore, this dataset is not limited to small objects. The medium and large objects account for nearly 29% of the dataset instances. The dataset also covers acquisition conditions at various heights and angles. In addition to vertical views, the data includes multiple oblique views captured at flight altitudes ranging from 60 to 120 meters. drone; multimodal; object detection; infrared-visible Liu Yu EMBARGO https://creativecommons.org/licenses/by-nc-nd/4.0/ dataset Science Data Bank