RF-DETR is a real-time transformer model architecture from Roboflow for object detection, instance segmentation, and keypoint detection (in preview). It is built on a DINOv2 vision transformer backbone, extends LW-DETR and Deformable DETR, and exposes all three tasks through one Python API. It is designed to be fine-tuned on custom datasets.
The published model sizes, from Nano to 2XLarge, were found with neural architecture search (NAS), and the same search is offered on the Roboflow platform to find an architecture for a specific dataset. Benchmarks are reported on Microsoft COCO and RF100-VL, with latency measured on an NVIDIA T4 using TensorRT, FP16, and batch size 1, alongside YOLO11, YOLO26, and LW-DETR baselines measured the same way.
Detection results for the RF-DETR sizes:
| Model | COCO AP50:95 | RF100-VL AP50:95 | Latency (ms) | Params (M) | Resolution |
|---|---|---|---|---|---|
| RF-DETR-N | 48.4 | 57.7 | 2.3 | 30.5 | 384x384 |
| RF-DETR-S | 53.0 | 60.2 | 3.5 | 32.1 | 512x512 |
| RF-DETR-M | 54.7 | 61.2 | 4.4 | 33.7 | 576x576 |
| RF-DETR-L | 56.5 | 62.2 | 6.8 | 33.9 | 704x704 |
| RF-DETR-XL | 58.6 | 62.9 | 11.5 | 126.4 | 700x700 |
| RF-DETR-2XL | 60.1 | 63.2 | 17.2 | 126.9 | 880x880 |
Segmentation models range from 40.3 to 48.8 COCO mask AP50:95 at 3.4 to 13.5 ms, and the keypoint preview model reaches 71.8 COCO keypoint AP50:95 at 9.7 ms. The XL and 2XL detection models ship as a separate Plus extension. Models can be trained in Google Colab or on the Roboflow platform, run directly from the rfdetr package or through Roboflow Inference, and annotated with Supervision.
Features
- Three vision tasks: object detection, instance segmentation, and keypoint detection (preview) behind one consistent API
- Model sizes: Nano through 2XLarge detection models and Nano through XLarge segmentation models, trading accuracy against latency
- DINOv2 backbone: a vision transformer backbone suited to fine-tuning on custom data
- Fine-tuning: train detection, segmentation, and keypoint models on your own datasets in Colab or on the Roboflow platform
- Neural architecture search: model sizes derived by NAS, with the same search available on the Roboflow platform per dataset
- Reproducible benchmarks: COCO and RF100-VL results measured in-house on a T4 with TensorRT FP16 against YOLO and LW-DETR baselines
- Inference options: predict from the Python package or through the Roboflow Inference library, with Supervision for annotation
