Synthetic Data for Data-Efficient Agricultural Computer Vision: Three Datasets and a Multi-Task Benchmark Across Real and Simulated Domains

This dataset release contains three agricultural computer vision datasets comprising both real-world and synthetically generated images: Apples: Fruit detection and instance segmentation in orchard environments Carrot Crack: Fine-grained defect detection and instance segmentation of surface cracks on carrots Potato–Stone: Separation of potatoes and stones in cluttered industrial sorting scenarios The datasets support research in object detection, instance segmentation, synthetic-to-real transfer learning, domain adaptation, data-efficient learning, and agricultural robotics. Dataset Summary Dataset Real Images Synthetic Images Real Instances Synthetic Instances Apples 1,047 1,000 40,728 49,329 Carrot Crack 110 1,000 568 24,861 Potato–Stone 50 1,000 2,028 70,544 Total dataset size 1,207 real images 3,000 synthetic images 43,324 annotated real instances 144,734 annotated synthetic instances Note: The real apple images in the training and validation sets were sourced from the MinneApple benchmark dataset for apple detection and segmentation (Häni, Roy, & Isler, 2019), as described in A Comparative Study of Fruit Detection and Counting Methods for Yield Mapping in Apple Orchards (Journal of Field Robotics, 2019). Data Format Images are provided in PNG or JPG format with corresponding instance segmentation annotations. Annotations follow a YOLO-compatible polygon format: class_id x1 y1 x2 y2 ... xn yn where polygon vertices are normalized to image dimensions. Class definitions: Apples: 0 = apple Carrot Crack: 0 = crack Potato–Stone: 0 = potato 1 = stone Synthetic datasets additionally include pixel-level instance masks. Research Applications This dataset can be used for: Synthetic-to-real transfer learning Domain adaptation Instance segmentation Object detection Data-efficient learning Few-shot learning Agricultural robotics Automated harvesting Produce quality inspection Industrial sorting and grading systems Dataset Characteristics The three datasets provide complementary challenges: Apples: Dense fruit detection with substantial occlusion and viewpoint variation. Carrot Crack: Fine-grained defect segmentation with limited real-world training data. Potato–Stone: Highly cluttered scenes with visually similar object categories and complex boundaries. Together, they provide a benchmark spanning sparse defect detection to highly crowded agricultural scenes. Associated Publication This dataset accompanies the study: Synthetic Data for Data-Efficient Agricultural Computer Vision: Three Datasets and a Multi-Task Benchmark Across Real and Simulated Domains License This dataset is released under the Attribution-NonCommercial-ShareAlike 3.0 United States. Funding This work was supported by: NORM.AI_sbo: Natural Objects Rendering for Economic AI ModelsProject Number: 2022-0578Flanders Make, Belgium Contact Wenzhi LiaoFlanders Make, [e-mail]