Large-scale orange fruit dataset for localization, classification and ripening assessment under varying environments

The presented dataset is a large-scale on-field orange image collection designed for fruit detection, classification, and ripening assessment in real-world orchard environments. It includes 5025 images captured under diverse weather, lighting conditions, and ripening stages. Images were collected using different smartphone cameras and preprocessed through a custom cropping algorithm to optimize annotation efficiency. The dataset was labeled using a semi-automated approach, combining YOLO-based pre-annotations refined manually in Roboflow. Annotation quality was further improved through a CLIP-based verification process to filter incorrect labels. The dataset is released with both YOLO and COCO annotations, enabling compatibility with multiple object detection frameworks. Additionally, a benchmark evaluation was conducted using state-of-the-art models, including YOLO (v5, v8, v10, v11) and RT-DETR, assessed via standard precision, recall, and F1-score metrics. Results showed that recent YOLO models (YOLOv10 and YOLOv11) achieved high detection performance, with mAP@0.5 values close to 0.892, and consistently outperform RT-DETR baselines (mAP@0.5 = 0.851) in terms of precision and inference speed. In this context, the structured design and high environmental diversity of the proposed dataset make it a valuable resource for developing and evaluating computer vision solutions in precision agriculture, including fruit ripening assessment, yield estimation, and automated harvesting.