Skip to content

About

An extension of RF-DETR that unlocks larger, more powerful detection models for when you need the highest possible accuracy.

Topics

Resources

Contributing

Stars

65 stars

Watchers

1 watching

Forks

Repository files navigation

RF-DETR+: Platform Detection Models for RF-DETR

version downloads python-version license

discord Papers with Code leaderboard

RF-DETR is the core package in the ecosystem. It provides the full training and inference stack, the {Nano, Small, Medium, Large} model lineup, and the APIs most users build on. RF-DETR+ is an extension package for RF-DETR that adds the XLarge and 2XLarge detection models for maximum accuracy and the Atto, Femto and Pico detection models for the lowest latency.

RF-DETR+ models use a DINOv2 vision transformer backbone at higher resolutions and larger feature dimensions than the core RF-DETR lineup, pushing state-of-the-art accuracy on Microsoft COCO and RF100-VL while retaining real-time inference speeds. Atto, Femto and Pico instead use Meta's Perception Encoder PE-Core-T backbone, with architectures selected by neural architecture search, to run in 1-2 ms on a T4. Use RF-DETR for the standard model set and RF-DETR+ when you need the highest-accuracy or the fastest variants.

Install

Install RF-DETR+ in a Python>=3.10 environment with pip. This will also install rfdetr as a dependency, which provides the core APIs and model definitions.

pip install rfdetr-plus
Install from source
pip install git+https://github.com/roboflow/rf-detr-plus.git

Benchmarks

RF-DETR+ XLarge and 2XLarge sit at the top of the RF-DETR accuracy/latency curve, delivering the highest COCO AP scores in the family, while Atto, Femto and Pico extend it below Nano. All latency numbers were measured on an NVIDIA T4 using TensorRT, FP16, and batch size 1.

Size Class COCO AP50 COCO AP50:95 RF100VL AP50 RF100VL AP50:95 Latency (ms) Params (M) Resolution Package / License
⭐ A RFDETRAtto 49.4 30.5 78.2 48.3 1.0 7.4 380x380 rfdetr_plus / PML 1.0
⭐ F RFDETRFemto 55.9 37.8 82.5 53.8 1.4 7.9 384x384 rfdetr_plus / PML 1.0
⭐ P RFDETRPico 60.2 41.6 84.1 56.0 1.7 8.4 560x560 rfdetr_plus / PML 1.0
N RFDETRNano 67.6 48.4 85.0 57.7 2.3 30.5 384x384 rfdetr / Apache 2.0
S RFDETRSmall 72.1 53.0 86.7 60.2 3.5 32.1 512x512 rfdetr / Apache 2.0
M RFDETRMedium 73.6 54.7 87.4 61.2 4.4 33.7 576x576 rfdetr / Apache 2.0
L RFDETRLarge 75.1 56.5 88.2 62.2 6.8 33.9 704x704 rfdetr / Apache 2.0
⭐ XL RFDETRXLarge 77.4 58.6 88.5 62.9 11.5 126.4 700x700 rfdetr_plus / PML 1.0
⭐ 2XL RFDETR2XLarge 78.5 60.1 89.0 63.2 17.2 126.9 880x880 rfdetr_plus / PML 1.0

Run Models

Install RF-DETR+ to use XL and 2XL models alongside the core RF-DETR lineup:

pip install rfdetr_plus

RF-DETR+ models require you to accept the Platform Model License before use. Once accepted, usage mirrors the standard RF-DETR API -- you import the models from rfdetr_plus and keep using the rfdetr utilities:

import requests
import supervision as sv
from PIL import Image
from rfdetr_plus import RFDETRXLarge
from rfdetr.assets.coco_classes import COCO_CLASSES

model = RFDETRXLarge(accept_platform_model_license=True)

image = Image.open(requests.get("https://media.roboflow.com/dog.jpg", stream=True).raw)
detections = model.predict(image, threshold=0.5)

labels = [f"{COCO_CLASSES[class_id]}" for class_id in detections.class_id]

annotated_image = sv.BoxAnnotator().annotate(image, detections)
annotated_image = sv.LabelAnnotator().annotate(annotated_image, detections, labels)

Train Models

RF-DETR+ models support fine-tuning with the same training API as core RF-DETR. You can train on your own dataset or use datasets from Roboflow Universe. Training needs the training extras of the core package, so install them first with pip install "rfdetr[train]".

from rfdetr_plus import RFDETRXLarge

model = RFDETRXLarge(accept_platform_model_license=True)
model.train(dataset_dir="path/to/dataset", epochs=50, lr=1e-4)

Export Atto, Femto or Pico for CPU

Install pip install "rfdetr-plus[onnx]". The compact ONNX export uses batch size one at the detector's configured resolution and targets the standard encoder and decoder layout for each size. It derives its weights from the loaded model and computes activation scales at inference time.

from rfdetr import RFDETRInference, from_checkpoint
from rfdetr_plus import RFDETRAtto

# The released COCO Atto fits this optional 8 MB storage budget.
pretrained = RFDETRAtto(device="cpu")
pretrained.export_cpu_onnx("output/pretrained", size_limit_bytes=8_000_000)

trained = from_checkpoint("checkpoint_best_total.pth", device="cpu")
artifact = trained.export_cpu_onnx("output/finetuned")
predictions = RFDETRInference(artifact, device="cpu").predict("image.jpg")

The optional size_limit_bytes checks the complete uncompressed file before publishing the artifact. The call raises ValueError if the file reaches or exceeds that limit. The reference Femto and Pico CPU exports are 11.77 MB and 13.10 MB respectively.

The encoder uses fused rotary and attention operators. Every size quantizes the 36 encoder projections (qkv, proj and fc1 of all twelve blocks) to dynamic INT8, and each size adds its own recipe on top:

  • Atto: fc2 of encoder blocks 6-11 also uses dynamic INT8, the 1x1 projector convolutions become float NHWC matrix multiplications, and the remaining large filters are stored as scaled INT8.
  • Femto: linear2 of decoder layers.1 also uses dynamic INT8, and the remaining large filters are stored in FP16.
  • Pico: linear1 of decoder layers.1 also uses dynamic INT8, the 1x1 projector convolutions become float NHWC matrix multiplications, and the remaining large filters are stored in FP16.

The export fails before tracing for any other size, for a non-PE-Core-T encoder, and for Femto or Pico with fewer than two decoder layers. The file uses ONNX opset 23, so run it with onnxruntime>=1.23.2, the tested floor.

This differs from the base model.export(format="onnx", quantization="int8"), which writes a static QDQ model with INT8 weights and activations and requires calibration_data. The compact export needs no calibration data. output_name follows the base full-stem semantics: the file is {output_name}.onnx, and the default is the model's size name plus -cpu, such as rfdetr-atto-cpu.onnx. Compare the compact export with model.export(format="onnx") on your validation images and target CPU to measure the accuracy and latency trade-off.

Documentation

Visit the RF-DETR documentation website to learn more about training, export, deployment, and the full model lineup.

License

RF-DETR+ code and model checkpoints are licensed under the Platform Model License 1.0 (PML-1.0). See LICENSE for details. These models require a Roboflow account to run and fine-tune.

The core RF-DETR models (Nano through Large) are available under the Apache License 2.0 in the rfdetr package.

Acknowledgements

Our work is built upon LW-DETR, DINOv2, and Deformable DETR. Thanks to their authors for their excellent work!

Citation

If you find our work helpful for your research, please consider citing the following BibTeX entry.

@misc{rf-detr,
    title={RF-DETR: Neural Architecture Search for Real-Time Detection Transformers},
    author={Isaac Robinson and Peter Robicheaux and Matvei Popov and Deva Ramanan and Neehar Peri},
    year={2025},
    eprint={2511.09554},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2511.09554},
}

Contribute

We welcome and appreciate all contributions! If you notice any issues or bugs, have questions, or would like to suggest new features, please open an issue or pull request. By sharing your ideas and improvements, you help make RF-DETR better for everyone.

About

An extension of RF-DETR that unlocks larger, more powerful detection models for when you need the highest possible accuracy.

Topics

Resources

Contributing

Stars

65 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages