Overview The ChestnutDetDataset was developed to support the development and evaluation of computer vision models for detecting on-ground chestnuts, with the ultimate goal of enabling automated chestnut picking technologies in agricultural robotics. The dataset contains high-resolution orchard ground images with manually annotated chestnut instances, enabling benchmarking of real-time object detection models. Dataset Contents The dataset is organized into two folders: ChestnutDetDataset/│├── Imagery/ # Raw RGB images (.JPG)└── Annotations/ # Annotation files (.JSON) Images: 319 high-resolution RGB images Resolution: 4032 × 3024 pixels Format: JPG Annotations: JSON files (same filename as corresponding images) Each image file corresponds directly to an annotation file with the same file name. Data Collection The imagery was collected in a commercial chestnut orchard (Owosso, Michigan, USA) during the 2024 chestnut harvest season. Images were acquired using a handheld smartphone during a morning orchard visit. Data collection involved walking through the orchard and recording ground scenes, resulting in diverse visual conditions, including: varying grass coverage different soil backgrounds naturally occurring lighting variations All images follow a consistent and self-explanatory naming convention. Annotations Annotations were created using the VGG Image Annotator (VIA). Trained personnel manually labeled exposed chestnuts (not including those hidden in the unopened burrs) in each image by drawing bounding boxes for individual chestnut instances. Bounding boxes follow the COCO format: [x, y, width, height], where: x, y – coordinates of the top-left corner of the bounding box, width – bounding box width, and height – bounding box height To ensure annotation quality, the initial annotations were subjected to separate review and correction. In total, the dataset contains: 319 images 6,524 annotated chestnut instances Benchmark Study A benchmark evaluation using this dataset was conducted on a range of real-time object detection models, including: 14 models from the YOLO (v11–v13) family 15 models from the RT-DETR (v1–v4) family Multiple model scales were tested to assess detection performance in terms of detection accuracy, inference times, and model complexity. Full details of the dataset, modeling methodology, and experiments are described in the associated publication: Fang, K., Lu, Y., Mu, X., 2026. Detection of On-Ground Chestnuts Using Artificial Intelligence Toward Automated Picking. AgriEngineering. The software programs and models developed in this study are available at https://github.com/AgFood-Sensing-and-Intelligence-Lab/ChestnutDetection Citation If you use this dataset in published research, please consider citing the dataset and/or the associated journal article. Contact We hope ChestnutDetDataset contributes to advancing research in agricultural robotics and automated harvesting systems. Please contact luyuzhen@msu.edu for questions, comments, and suggestions regarding the dataset.