dronefreak/hrp4k-rfdetr-medium

RF-DETR Medium Finetuned on HRP4K

Fine-tuned RF-DETR Medium object detector on the HRP4K 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	55.66
mAP@50-95	31.81
Precision	64.25
Recall	51.9
F1 Score	57.42
Parameters	33.7M
FLOPs	N/A (not published upstream)

---

Evaluation Protocol

Metrics reported in this model card are computed on the HRP4K test split, using DetectionBench's standard evaluation pipeline (detectionbench-evaluate).

---

HRP4K Model Zoo

Every model DetectionBench has trained and evaluated on HRP4K 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
RF-DETR Small	56.04	31.54	62.01	55.48
RF-DETR Medium	55.66	31.81	64.25	51.9
RF-DETR Nano	52.06	28.62	62.03	47.88
YOLOv26m	51.62	29.83	57.33	49.89
YOLOv26s	48.65	27.16	60.04	46.91
YOLOv8n	48.58	26.24	58.0	46.58
YOLOv8s	47.59	26.71	59.81	44.3
YOLOv8m	47.26	26.77	59.56	44.52

---

Per-Class Performance

Class	mAP@50	mAP@50-95
pothole	55.66	31.81

---

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 HRP4K. For the full dataset description, provenance, license, and citation, see the dataset card:

https://huggingface.co/datasets/dronefreak/HRP4K

Classes

• pothole

---

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/hrp4k-rfdetr-medium",
    filename="checkpoint_best_total.pth"
)

model = rfdetr.RFDETRMedium(pretrain_weights=weights)

Run Inference

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

---

Training Configuration

Setting	Value
Dataset	HRP4K
Framework	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	100
Epochs (actually trained)	21
Early Stopping Patience	15
Batch Size	2
Resolution	1024
Optimizer	adamw
Learning Rate	0.0001
Seed	42

---

Repository Contents

checkpoint_best_total.pth
metrics.csv
config.json
hrp4k_rfdetr-medium_showcase.jpg
README.md

---

Related Resources

• HRP4K dataset card on Hugging Face
• DetectionBench -- reproducible benchmarks for modern object detectors on real-world datasets
• HRP4K paper (Scientific Data, doi:10.1038/s41597-026-07317-w)
• HRP4K dataset record (Zenodo, doi:10.5281/zenodo.17522874)
• HRP4K project repository

---

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

• Incomplete upstream train split: the official train.json references 4,203 images, but the released archive ships only 2,286 of them, so the training set here is 2,286 images with 2,790 boxes (the validation and test splits are complete at 900 images each).
• Single class, sparse targets: every box is a pothole, with only about 1.2 boxes per image, and roughly a third of the images in every split contain no pothole at all (789 of 2,286 train, 300 of 900 valid, 300 of 900 test). Those empty test frames mean a false positive on a clean road directly lowers precision.
• Small targets in very large frames: every image is 3840x2160 (4K), and the median pothole is only about 61 px wide on the native frame (0.05% of the image area; the smallest 5% are under 16 px), so results depend strongly on the input resolution the model was trained and evaluated at -- at 640 px most potholes shrink to a few pixels.
• Wide, perspective-view boxes: the median box is about 2.9 times wider than tall, reflecting the dataset's road-surface perspective, so generalization to other camera positions (e.g. aerial or vehicle mounted) is untested.
• Small test split: 900 images with 921 boxes, so differences of about one mAP point between models are within noise and should be read with caution.

---

Citation

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

1. The HRP4K dataset (see below)
2. The original RF-DETR Medium 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{chen2026hrp4k,
  title={A high-resolution perspective-view road image dataset for pothole detection},
  author={Chen, Hanshen and Tu, Zhoulin and Zhao, Yu and Ye, Jianfeng},
  journal={Scientific Data},
  volume={13},
  pages={961},
  year={2026},
  doi={10.1038/s41597-026-07317-w}
}

@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 HRP4K in this model card:

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}
}