GWHD 2021: Global Wheat Head Dataset (Object Detection)

Unofficial redistribution of the Global Wheat Head Dataset (GWHD) 2021 competition release, reformatted into a standardized YOLO-compatible directory layout.

Disclaimer

This repository is not an official release of the Global Wheat Head Dataset.

GWHD was created by the Global Wheat Head Detection consortium — a multi-institution, multi-country collaboration (David, Serouart, Madec, et al.) — who retain all copyright and intellectual property rights. This repository does not claim ownership of any images, annotations, or metadata.

This repository exists for two purposes:

1. To reorganize the dataset into a standardized YOLO/Ultralytics-compatible directory structure that can be used directly by many modern object detection training pipelines.
2. To provide a more reliable download source, as the original hosting may be slow or difficult to access.

---

Dataset Description

GWHD 2021 is a dense, single-class object detection benchmark for wheat head localization, assembled from field images captured across multiple countries and research institutions to maximize genotype, growth-stage, and imaging-condition diversity. The official release reports 6,422 images and 275,187 wheat head annotations; this repository's own direct measurement of the competition CSV splits it was built from (see below) found 6,515 images and 275,468 boxes — small discrepancies from the paper's headline figures that likely reflect minor differences between the full consortium release and the specific competition_{train,val,test}.csv files distributed for the 2021 competition.

This repository preserves the dataset's images and all annotated boxes while re-encoding the label format for YOLO compatibility (see Changes from the Official Release below).

---

Changes from the Official Release

Unlike a simple directory reorganization, this conversion involved a genuine annotation format change — the raw release does not ship YOLO-format labels. Specifically:

• Box encoding converted. The original release stores all boxes for an image as a single semicolon-separated BoxesString CSV cell, each box an absolute-pixel "x1 y1 x2 y2" (top-left/bottom-right) tuple. This repository converts every box into YOLO's normalized class x_center y_center width height format, one line per box, in a .txt file per image. The underlying coordinates are a straightforward affine transform of the originals — no boxes were resized, cropped, or otherwise altered in content, only re-expressed.
• Images with zero wheat heads preserved. The original CSV uses a "no_box" sentinel for images with no annotated wheat heads (~2% of images); this repository preserves those images with an empty label file rather than dropping them.
• A small upstream data quirk was fixed. A handful of rows in the original CSVs (2 in train, 1 in test) reuse the same image filename with different domain and box annotations attached — most likely a labeling/export artifact in the original release, not something introduced here. Left as-is, the second occurrence would silently overwrite the first during YOLO conversion. This repository disambiguates such rows (renaming the second occurrence's copy, e.g. image__1.png) so no annotations from the original release were dropped.
• No image pixel content was modified.
• No splits were changed; train/val/test map directly to the original competition CSVs of the same name (val renamed to valid to match this repository's directory convention).

---

Dataset Structure

dataset/
├── README.md
├── data.yaml
├── train/
│   ├── images/
│   └── labels/
├── valid/
│   ├── images/
│   └── labels/
└── test/
    ├── images/
    └── labels/

where:

• images/ contains the original 1024x1024 RGB field-imagery PNGs for each split.
• labels/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized; empty for images with no wheat heads).
• data.yaml is the Ultralytics dataset configuration file (class names, split paths).
• Splits (as measured in this repository): train 3,657 images / 163,690 boxes · valid 1,476 images / 44,347 boxes · test 1,382 images / 67,431 boxes (6,515 images / 275,468 boxes total).

Classes (1)

wheat_head — this is a single-class, dense-detection dataset; images can contain from zero to 100+ annotated wheat heads.

---

Dataset Sources

Papers

Global Wheat Head Dataset 2021: more diversity to improve the benchmarking of wheat head localization methods

David, Serouart, Smith, Madec, Velumani, Liu, Wang, Pinto Espinosa, Shafiee, Tahir, Tsujimoto, Nasuda, Zheng, Kichgessner, Aasen, Hund, Sadhegi-Tehran, Nagasawa, Ishikawa, Dandrifosse, Carlier, Mercatoris, Kuroki, Wang, Ishii, Badhon, Pozniak, LeBauer, Lilimo, Poland, Chapman, de Solan, Baret, Stavness, Guo

Plant Phenomics, 2021. DOI: 10.34133/2021/9846158 (preprint: arXiv:2105.07660)

This 2021 release extends the original:

Global Wheat Head Detection (GWHD) Dataset: A Large and Diverse Dataset of High-Resolution RGB-Labelled Images to Develop and Benchmark Wheat Head Detection Methods

David et al., Plant Phenomics, 2020. DOI: 10.34133/2020/3521852

Official Resources

• Official Website: https://www.global-wheat.com/
• Official Hugging Face Organization: https://huggingface.co/GlobalWheat (hosts related consortium datasets; this is a separate, unofficial redistribution of the 2021 object-detection competition release specifically)

---

Attribution

All credit for the dataset belongs entirely to the Global Wheat Head Detection consortium and paper authors listed above.

This repository only redistributes their data with a YOLO-format label conversion, for improved usability and accessibility.

If you use this dataset in your research, please cite the original publications below.

---

License

The Global Wheat Head Dataset 2021 is distributed by its creators under Creative Commons Attribution 4.0 International (CC BY 4.0), as stated in the dataset paper. The dataset's official website additionally describes the data as "free to use for any purpose."

Accordingly:

• Attribution to the original creators is required.
• Commercial and non-commercial use are both permitted.
• No share-alike obligation (derivative works are not required to use the same license).

This repository is distributed under the same license.

---

Citation

If you use this dataset, please cite both papers:

@article{david2021global,
  title = {Global Wheat Head Dataset 2021: more diversity to improve the benchmarking of wheat head localization methods},
  author = {David, Etienne and Serouart, Mario and Smith, Daniel and Madec, Simon and Velumani, Kaaviya and Liu, Shouyang and Wang, Xu and Pinto Espinosa, Francisco and Shafiee, Shahameh and Tahir, Izzat S. A. and Tsujimoto, Hisashi and Nasuda, Shuhei and Zheng, Bangyou and Kichgessner, Norbert and Aasen, Helge and Hund, Andreas and Sadhegi-Tehran, Pouria and Nagasawa, Koichi and Ishikawa, Goro and Dandrifosse, S{\'e}bastien and Carlier, Alexis and Mercatoris, Benoit and Kuroki, Ken and Wang, Haozhou and Ishii, Masanori and Badhon, Minhajul A. and Pozniak, Curtis and LeBauer, David Shaner and Lilimo, Morten and Poland, Jesse and Chapman, Scott and de Solan, Benoit and Baret, Fr{\'e}d{\'e}ric and Stavness, Ian and Guo, Wei},
  journal = {Plant Phenomics},
  year = {2021},
  doi = {10.34133/2021/9846158}
}

@article{david2020global,
  title = {Global Wheat Head Detection (GWHD) Dataset: A Large and Diverse Dataset of High-Resolution RGB-Labelled Images to Develop and Benchmark Wheat Head Detection Methods},
  author = {David, Etienne and Madec, Simon and Sadeghi-Tehran, Pouria and Aasen, Helge and Zheng, Bangyou and Liu, Shouyang and Kirchgessner, Norbert and Ishikawa, Goro and Nagasawa, Koichi and Badhon, Minhajul A. and Pozniak, Curtis and de Solan, Benoit and Hund, Andreas and Chapman, Scott C. and Baret, Fr{\'e}d{\'e}ric and Stavness, Ian and Guo, Wei},
  journal = {Plant Phenomics},
  year = {2020},
  doi = {10.34133/2020/3521852}
}

---

Acknowledgements

We sincerely thank the Global Wheat Head Detection consortium and all contributing institutions for assembling, annotating, and publicly releasing this valuable precision-agriculture benchmark.