RF-DETR Medium Finetuned on ExDark

Fine-tuned RF-DETR Medium object detector on the ExDark 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.

ExDark Detection Demo

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

model = rfdetr.RFDETRMedium(pretrain_weights=weights)

Run Inference

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

Performance

Evaluated on the ExDark test split, using DetectionBench's standard evaluation pipeline (detectionbench-evaluate).

Metric	Score (%)
mAP@50	88.64
mAP@50-95	62.55
Precision	86.6
Recall	79.46
F1 Score	82.88
Parameters	33.7M
FLOPs	N/A (not published upstream)

ExDark Model Zoo

Every model DetectionBench has trained and evaluated on ExDark 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	88.98	61.67	83.07	81.89
RF-DETR Medium	88.64	62.55	86.6	79.46
RF-DETR Nano	85.27	58.01	85.18	74.67
YOLO26l	77.51	50.88	80.71	70.72
YOLO26m	76.54	50.02	82.29	68.83
YOLOv8x	75.4	48.39	81.53	65.86
YOLOv8l	75.26	48.48	81.44	67.58
YOLOv8m	74.69	48.05	78.4	69.17
YOLO11x	74.41	48.98	81.87	67.05
YOLOv9m	74.17	47.38	76.27	67.94
YOLO26s	74.0	48.32	79.11	65.59
YOLO11l	73.44	47.56	78.57	67.09
YOLO11s	73.35	46.8	77.93	66.38
YOLO11m	73.17	47.16	74.83	67.23
YOLOv8s	73.01	45.85	78.26	65.13
YOLO26n	72.7	46.27	81.0	62.67
YOLOv8n	71.29	44.78	78.25	62.76
YOLO11n	70.36	44.72	76.18	61.15

Per-Class Performance

Class	mAP@50	mAP@50-95
Bicycle	84.51	58.56
Boat	89.93	55.03
Bottle	81.39	54.66
Bus	92.25	75.09
Car	91.94	66.21
Cat	91.27	66.74
Chair	84.52	60.12
Cup	88.85	60.17
Dog	91.77	70.9
Motorbike	91.55	64.08
People	89.01	56.93
Table	86.74	62.17

This model was evaluated with Supervision's detection metrics, which report mAP/Precision/Recall directly but don't produce a confusion-matrix plot the way Ultralytics' validator does.

Dataset

This model was trained on ExDark. For the full dataset description, provenance, license, and citation, see the dataset card:

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

Classes

Bicycle
Boat
Bottle
Bus
Car
Cat
Chair
Cup
Dog
Motorbike
People
Table

Training Configuration

Setting	Value
Dataset	ExDark
Framework	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	500
Epochs (actually trained)	104
Early Stopping Patience	100
Batch Size	9
Resolution	576
Optimizer	adamw
Learning Rate	0.0001
Seed	42

Repository Contents

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

Related Resources

ExDark dataset card on Hugging Face
DetectionBench -- reproducible benchmarks for modern object detectors on real-world datasets

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

Severe class imbalance: People accounts for roughly 46% of all annotated boxes while Bus is the rarest class, so per-class accuracy on rare classes is measured on very few test examples and should be read with wide uncertainty.
Small dataset overall (7,344 images, 734 in the test split, across 12 classes) -- limited training signal for several classes independent of the imbalance above.
Two-hop provenance: this dataset was converted to YOLO format by a third-party Roboflow export before reaching DetectionBench, not sourced directly from the original per-class-folder release; images are pre-resized to 640x640 by that export.
The original authors separately request non-commercial use of this dataset (beyond the BSD-3-Clause license text itself) -- this applies to any model trained on it, not only the raw images.

Citation

If you use this model in your research, please consider citing the dataset and the model architecture:

@article{Exdark,
  title = {Getting to Know Low-light Images with The Exclusively Dark Dataset},
  author = {Loh, Yuen Peng and Chan, Chee Seng},
  journal = {Computer Vision and Image Understanding},
  volume = {178},
  pages = {30-42},
  year = {2019},
  doi = {https://doi.org/10.1016/j.cviu.2018.10.010}
}

@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}
}