SyneticAI/ApplesM5-Dataset

🍎 ApplesM5: Synthetic Apple Detection Benchmark

This repository hosts the data files (images and annotations) used in the Synetic AI research paper, "Better Than Real: Synthetic Apple Detection for Orchards." This dataset was created through procedural content generation and physically-based rendering (PBR) to provide a clean, highly generalized training signal for robust agricultural AI.

The data demonstrates that training exclusively on this synthetic dataset yields superior generalization compared to models trained solely on real-world data, achieving up to a +34.24% increase in mAP50-95.

Dataset Structure and Format

The dataset is provided in a file-based structure optimized for training YOLO models.

Split	Description	Format	Total File Count
train/	Synthetic, procedurally generated images and labels. (Used for training.)	YOLOv8 (1 class)	> 10,000
val/	Real-world image samples from external orchards. (Used for validation/testing.)	YOLOv8 (1 class)	~300

Citation

Please cite the associated whitepaper when using this dataset in your research:

@article{synetic2025applesm5,
  title={{Better Than Real: Synthetic Apple Detection for Orchards}},
  author={Blaga, Octavian and Scott, David and Zand, Ramtin and Seekings, James Blake},
  journal={ResearchGate preprint},
  year={2025},
  doi={10.13140/RG.2.2.29696.49920},
  url={https://www.researchgate.net/publication/397341880_Better_Than_Real_Synthetic_Apple_Detection_for_Orchards},
  note={Code available at: \url{https://github.com/Syneticai/ApplesM5}}
}