AIRST-CS

AIRST-CS, short for Aerial Infrared–RGB Small Targets: Cars and Ships, is a paired visible–infrared dataset for UAV-based multispectral small-object detection. The research hypothesis is that visible and infrared aerial images provide complementary cues for detecting small cars and ships in complex monitoring scenarios. Visible images preserve texture, color, and structural details, while infrared images provide thermal information that can remain useful under low-light or nighttime conditions. The dataset contains 2,463 registered RGB–infrared image pairs with a resolution of 640 × 512 pixels. It is divided into 1,724 training pairs, 369 validation pairs, and 370 test pairs using a scene-level split. The data were collected using a DJI Matrice 300 UAV equipped with a Zenmuse H20N payload. The two modalities were hardware-synchronized, followed by calibration, cross-modal alignment, quality control, and manual annotation. Each image pair is annotated with axis-aligned bounding boxes for two categories: car and ship. Many targets occupy less than 0.5% of the image area, making the dataset suitable for small-object and tiny-object detection research. AIRST-CS covers urban and waterway scenes, including highways, ports, docks, rivers, and inland waterways, with daytime/nighttime imaging, strong illumination variation, cluttered backgrounds, dense targets, occlusion, and thermal artifacts. The dataset can be used for RGB–infrared object detection, multispectral fusion, modality reliability analysis, robust cross-modal learning, and UAV-based visual perception. Users should follow the official train/validation/test split for reproducible evaluation.