dronefreak/gwhd-rfdetr-small

RF-DETR Small Finetuned on Global Wheat Head Dataset

Fine-tuned RF-DETR Small object detector on the Global Wheat Head Dataset benchmark dataset, trained and evaluated as part of DetectionBench -- a framework for reproducibly benchmarking modern object detectors with identical training recipes and evaluation metrics across multiple real-world datasets.

---

Detection Showcase

---

Performance

Metric	Score (%)
mAP@50	64.51
mAP@50-95	26.12
Precision	77.6
Recall	63.61
F1 Score	69.91
Parameters	32.1M
FLOPs	N/A (not published upstream)

---

Evaluation Protocol

Metrics reported in this model card are computed on the Global Wheat Head Dataset test split, using DetectionBench's standard evaluation pipeline (detectionbench-evaluate).

---

Global Wheat Head Dataset Model Zoo

Every model DetectionBench has trained and evaluated on Global Wheat Head Dataset so far, for full transparency -- see DetectionBench for the smaller, curated comparison set used on the project README.

Model	mAP@50	mAP@50-95	Precision	Recall
YOLOv11x	74.25	34.92	83.37	67.92
YOLOv26m	71.58	33.48	79.94	62.74
YOLOv26s	70.49	31.39	79.47	63.43
YOLOv8m	69.55	29.34	80.69	63.74
YOLOv8s	68.29	29.48	79.89	62.52
RF-DETR Medium	67.1	27.11	78.6	65.66
YOLOv8n	66.19	28.21	78.69	60.3
RF-DETR Small	64.51	26.12	77.6	63.61
RF-DETR Nano	53.82	19.65	72.52	53.64

---

Per-Class Performance

Class	mAP@50	mAP@50-95
wheat_head	64.51	26.12

---

Evaluation Visualizations

This model was evaluated with Supervision's detection metrics, which report mAP/Precision/Recall directly but don't produce PR-curve, F1-curve, or confusion-matrix plot images the way Ultralytics' validator does. See the Performance table above for Precision/Recall/F1 and the per-class table above for the full per-class mAP breakdown.

---

Dataset

This model was trained on Global Wheat Head Dataset. For the full dataset description, provenance, license, and citation, see the dataset card:

https://huggingface.co/datasets/dronefreak/GWHD

Classes

• wheat_head

---

Usage

Install Dependencies

pip install rfdetr huggingface_hub

Load Model from Hugging Face

from huggingface_hub import hf_hub_download
import rfdetr

weights = hf_hub_download(
    repo_id="dronefreak/gwhd-rfdetr-small",
    filename="checkpoint_best_total.pth"
)

model = rfdetr.RFDETRSmall(pretrain_weights=weights)

Run Inference

detections = model.predict("image.jpg", threshold=0.25)

---

Training Configuration

Setting	Value
Dataset	Global Wheat Head Dataset
Framework	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	500
Epochs (actually trained)	124
Early Stopping Patience	100
Batch Size	6
Resolution	512
Optimizer	adamw
Learning Rate	0.0001
Seed	42

---

Repository Contents

checkpoint_best_total.pth
metrics.csv
config.json
gwhd_rfdetr-small_showcase.jpg
README.md

---

Related Resources

• Global Wheat Head Dataset dataset card on Hugging Face
• DetectionBench -- reproducible benchmarks for modern object detectors on real-world datasets
• GWHD 2021 paper preprint (arXiv:2105.07660)

---

Training Framework

This model was trained using DetectionBench, an open-source framework for benchmarking object detectors across multiple real-world datasets with a common pipeline.

Features include:

• A dataset-adapter registry for converting real-world datasets into a canonical format
• Identical training/evaluation recipes across model families (Ultralytics YOLO/RT-DETR, RF-DETR)
• Hardware profiling (latency, FPS, VRAM, parameters, FLOPs)
• One-command reproducibility via versioned Hydra configs

If you find this model useful, please consider starring the repository.

---

Known Limitations

• Single-class, dense-detection dataset: every box is wheat_head, so per-class breakdowns collapse to one row -- the challenge here is localization density, not class discrimination.
• Very dense annotations: ~45 wheat heads per image on average (up to 129 in a single image), with roughly 54% of boxes covering under 0.5% of the image area -- heavy box overlap and occlusion between adjacent heads is common.
• A small fraction of images (~1.4% of the training split) contain zero wheat heads and are still included with empty ground truth, matching the official release's convention.
• Assembled from field images across multiple countries/institutions with varying genotypes, growth stages, and imaging conditions -- performance on wheat varieties or field conditions not represented in this mix is untested.
• Distributed under CC BY 4.0 (attribution required, commercial use permitted) -- see the dataset card for the full license text and provenance.

---

Citation

If you use this model in your research, please consider citing:

1. The Global Wheat Head Dataset dataset (see below)
2. The original RF-DETR Small architecture (see below)
3. The other model architectures shown in the Model Zoo/External Comparison tables above, if you reference their results
4. DetectionBench, the training/evaluation framework used to produce this checkpoint

@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}
}

@inproceedings{robinson2026rfdetr,
  title     = {RF-DETR: Real-Time Detection Transformer},
  author    = {Robinson, Isaac and Robicheaux, Peter and Popov, Matvei and Ramanan, Deva and Peri, Neehar},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2511.09554}
}

@article{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and Darcet, Timoth{\'e}e and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and others},
  journal={arXiv preprint arXiv:2304.07193},
  year={2023}
}

Other architectures compared against on Global Wheat Head Dataset in this model card:

YOLOv11

No official YOLO11 research paper has been published by Ultralytics; the most commonly cited independent architectural analysis is used instead:

@article{khanam2024yolov11,
  title={YOLOv11: An Overview of the Key Architectural Enhancements},
  author={Khanam, Rahima and Hussain, Muhammad},
  journal={arXiv preprint arXiv:2410.17725},
  year={2024}
}

YOLOv26

@article{jocher2026yolo26,
  title={Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models},
  author={Jocher, Glenn and Qiu, Jing and Liu, Mengyu and Lyu, Shuai and Akyon, Fatih Cagatay and Kalfaoglu, Muhammet Esat},
  journal={arXiv preprint arXiv:2606.03748},
  year={2026}
}

YOLOv8

No official YOLOv8 research paper has been published by Ultralytics; this is their own recommended software citation instead:

@software{jocher2023yolov8,
  author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu},
  title = {Ultralytics YOLOv8},
  version = {8.0.0},
  year = {2023},
  url = {https://github.com/ultralytics/ultralytics},
  license = {AGPL-3.0}
}

@software{Saksena_DetectionBench_2026,
  author = {Saksena, Saumya Kumaar},
  title = {DetectionBench: Reproducible Benchmarks for Modern Object Detectors on Real-World Datasets},
  url = {https://github.com/dronefreak/DetectionBench},
  year = {2026}
}