Paired RGB–Near-Infrared Grape Dataset for Object Detection

This dataset contains 2,001 spatially aligned RGB and active near-infrared (NIR) image pairs acquired for grape object detection in agricultural environments. The images were captured using a Kinect RGB-D camera under natural daylight and nighttime low-light conditions. The dataset includes variations in illumination, background clutter, grape color, object scale, and occlusion by leaves and branches. The aligned observations have a resolution of 1280 × 720 pixels. A total of 8,343 grape instances are annotated in two categories: green grape and purple grape. The dataset is divided into 1,600 training pairs and 401 validation pairs. Green-grape and purple-grape instances account for 51.01% and 48.99% of all annotated instances, respectively. The annotations are provided in YOLO object-detection format. Each aligned RGB–NIR image pair shares the corresponding object annotation. The dataset was checked for missing modality pairs, filename inconsistencies, invalid coordinates, and duplicated bounding boxes. Seven duplicated bounding boxes identified during quality control were removed. Only the RGB and near-infrared modalities were used as model inputs in the associated study. This dataset supports research on RGB–NIR multimodal object detection, agricultural vision, multimodal feature fusion, and robustness to cross-modal misregistration or missing modalities.