RF-DETR Nano Finetuned on BDD100K

Fine-tuned RF-DETR Nano object detector on the BDD100K 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.

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/bdd100k-rfdetr-nano",
    filename="checkpoint_best_total.pth"
)

model = rfdetr.RFDETRNano(pretrain_weights=weights)

Run Inference

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

Performance

Evaluated on the BDD100K test split, using DetectionBench's standard evaluation pipeline (detectionbench-evaluate).

Metric	Score (%)
mAP@50	56.9
mAP@50-95	31.58
Precision	80.68
Recall	64.78
F1 Score	71.86
Parameters	30.5M
FLOPs	N/A (not published upstream)

BDD100K Model Zoo

Every model DetectionBench has trained and evaluated on BDD100K 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
YOLO26m	61.55	35.88	77.04	56.05
YOLOv10m	61.13	35.63	76.78	55.65
YOLO11m	61.1	35.57	66.78	55.61
YOLOv8m	60.89	35.47	66.52	55.42
YOLO26s	58.76	33.86	75.49	53.07
YOLOv8s	57.93	33.25	75.38	51.63
YOLOv10s	57.64	33.34	75.02	52.12
YOLO11s	57.63	33.1	74.31	52.42
RF-DETR Nano	56.9	31.58	80.68	64.78
YOLO26n	52.25	29.23	72.56	46.87
YOLOv9t	52.04	29.46	71.34	46.72
YOLOv10n	51.95	29.31	71.58	46.62
YOLOv8n	51.67	29.09	70.95	46.59
YOLO11n	51.63	29.06	71.68	46.34

Per-Class Performance

Class	mAP@50	mAP@50-95
person	61.5	29.93
rider	49.23	25.13
car	78.84	46.87
truck	66.7	47.33
bus	67.21	50.21
train	5.43	3.79
motor	52.81	26.78
bike	52.34	25.74
traffic light	64.87	23.88
traffic sign	70.08	36.18

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 BDD100K. For the full dataset description, provenance, license, and citation, see the dataset card:

https://huggingface.co/datasets/dronefreak/BDD100K

Classes

person
rider
car
truck
bus
train
motor
bike
traffic light
traffic sign

Training Configuration

Setting	Value
Dataset	BDD100K
Framework	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	30
Epochs (actually trained)	29
Early Stopping Patience	8
Batch Size	16
Resolution	576
Optimizer	adamw
Learning Rate	0.0001
Seed	42

Repository Contents

checkpoint_best_total.pth
metrics.csv
config.json
bdd100k_rfdetr-nano_showcase.jpg
assets/demo_banner.mp4
assets/demo_banner_poster.jpg
README.md

Related Resources

BDD100K dataset card on Hugging Face
DetectionBench -- reproducible benchmarks for modern object detectors on real-world datasets
BDD100K project website (official data source)

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

Not comparable to the official BDD100K test-server leaderboard: the official test split has no released labels, so the test split here is BDD100K's official validation set (10,000 images) and a seeded 15% slice of the official train set is held out for validation.
Severe class imbalance: car (55.4%), traffic sign (18.6%) and traffic light (14.5%) dominate the boxes, while rider (0.4%), motor (0.2%) and especially train (about 150 boxes in the whole dataset) are rare -- per-class accuracy on those classes is measured on very few examples and is close to noise for train.
Small objects: the median box covers only 0.09% of the 1280x720 frame, and traffic lights and signs are the smallest and hardest classes (medians of roughly 16 px and 21 px at native resolution), so scores on them depend heavily on input resolution.
Detection labels only: BDD100K's lane-marking and drivable-area annotations are dropped, so these models cover the 2D object detection task only.
Conditions are not broken down: the images span weather, time-of-day and scene conditions, but scores here are aggregated over all of them, and generalization outside the US road scenes BDD100K covers is untested.
Non-commercial redistribution only: BDD100K's own data license (reproduced in full on the dataset card) permits redistribution for educational, research and not-for-profit purposes with notice -- commercial use needs separate permission from UC Berkeley's Office of Technology Licensing, not a blanket no-redistribution grant.
Label vintage is 2018, not the paper's 2020 publication date: the detection labels mirrored here are BDD100K's original 2018 annotation release (matching the data license's own copyright year), carried through a two-hop Kaggle mirror -- see the dataset card for why a separate, later-dated label release couldn't be located or verified.

Citation

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

@inproceedings{yu2020bdd100k,
  title={BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning},
  author={Yu, Fisher and Chen, Haofeng and Wang, Xin and Xian, Wenqi and Chen, Yingying and Liu, Fangchen and Madhavan, Vashisht and Darrell, Trevor},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={2636--2645},
  year={2020}
}

@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}
}