dronefreak/lisa-rfdetr-medium

RF-DETR Medium Finetuned on LISA Traffic Lights

Fine-tuned RF-DETR Medium object detector on the LISA Traffic Lights 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	33.01
mAP@50-95	14.12
Precision	69.37
Recall	58.28
F1 Score	63.34
Parameters	33.7M
FLOPs	N/A (not published upstream)

---

Evaluation Protocol

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

---

LISA Traffic Lights Model Zoo

Every model DetectionBench has trained and evaluated on LISA Traffic Lights 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 Medium	33.01	14.12	69.37	58.28
RF-DETR Small	32.7	15.11	69.7	53.9
YOLOv26m	29.08	13.68	43.79	29.71
YOLOv26x	28.92	14.09	43.59	26.96
RF-DETR Nano	27.47	12.16	74.96	52.57
YOLOv26l	27.28	13.54	41.42	27.99
YOLOv26s	26.91	12.82	42.47	26.43
YOLOv11x	26.4	13.09	53.98	24.38
YOLOv8m	25.07	11.91	38.16	25.15
YOLOv26n	23.74	10.57	37.23	25.7

---

Per-Class Performance

Class	mAP@50	mAP@50-95
go	65.38	32.49
goForward	0.0	0.0
goLeft	16.93	5.68
stop	63.34	23.78
stopLeft	15.85	6.68
warning	53.7	25.46
warningLeft	15.88	4.73

---

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 LISA Traffic Lights. For the full dataset description, provenance, license, and citation, see the dataset card:

https://huggingface.co/datasets/dronefreak/LISA-Traffic-Lights

Classes

• go
• goForward
• goLeft
• stop
• stopLeft
• warning
• warningLeft

---

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/lisa-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	LISA Traffic Lights
Framework	RF-DETR
Training Toolkit	DetectionBench
Epochs (configured max)	500
Epochs (actually trained)	102
Early Stopping Patience	100
Batch Size	4
Resolution	576
Optimizer	adamw
Learning Rate	0.0001
Seed	42

---

Repository Contents

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

---

Related Resources

• LISA Traffic Lights 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

• Rare/underrepresented arrow classes (goForward, goLeft, stopLeft, warningLeft) have far fewer training examples than the base go/stop/warning classes and correspondingly lower detection accuracy across every model in this zoo.
• Sequential dashcam video frames mean visually similar consecutive frames can appear within the same split; performance on genuinely novel scenes may differ from the reported test-split numbers.
• Trained and evaluated only on San Diego daytime/nighttime driving sequences (Pacific Beach, La Jolla); generalization to different traffic-light hardware, road layouts, or camera setups is untested.
• Small, distant traffic lights are harder to detect reliably, consistent with general small-object detection challenges.

---

Citation

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

1. The LISA Traffic Lights 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{jensen2016vision,
  title={Vision for looking at traffic lights: Issues, survey, and perspectives},
  author={Jensen, Morten Born{\o} and Philipsen, Mark Philip and M{\o}gelmose, Andreas and Moeslund, Thomas Baltzer and Trivedi, Mohan Manubhai},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={17},
  number={7},
  pages={1800--1815},
  year={2016},
  doi={10.1109/TITS.2015.2509509},
  publisher={IEEE}
}

@inproceedings{philipsen2015traffic,
  title={Traffic light detection: A learning algorithm and evaluations on challenging dataset},
  author={Philipsen, Mark Philip and Jensen, Morten Born{\o} and M{\o}gelmose, Andreas and Moeslund, Thomas B and Trivedi, Mohan M},
  booktitle={Intelligent Transportation Systems (ITSC), 2015 IEEE 18th International Conference on},
  pages={2341--2345},
  year={2015},
  organization={IEEE}
}

@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 LISA Traffic Lights 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}
}