rf-detr

Overview of rf-detr

What it is

RF-DETR is a real-time transformer architecture for object detection, instance segmentation, and keypoint detection (preview) developed by Roboflow. It uses a DINOv2 vision transformer backbone (PE-Core-T backbone for Atto, Femto and Pico) and is built for fine-tuning. The architecture was presented at ICLR 2026.

Key features

• Real-time object detection, instance segmentation, and keypoint detection (preview) in a single API.
• Uses a DINOv2 vision transformer backbone with neural architecture search (NAS) for model sizing.
• Supports multiple model sizes: RF-DETR-N, -S, -M, -L (Apache 2.0) and RF-DETR-Atto, -Femto, -Pico, -XL, -2XL (PML 1.0).
• Achieves state-of-the-art accuracy and latency trade-offs on Microsoft COCO and RF100-VL benchmarks.
• Open-source rfdetr package and Apache-designated models released under Apache 2.0.
• Plus components (rfdetr_plus) and larger models are licensed under PML 1.0.

Use cases

• Fine-tuning custom object detection and segmentation models on user-provided datasets.
• Real-time inference applications that require low latency (e.g., edge devices, robotics).
• Instance segmentation and keypoint detection tasks where transformer-based architectures are preferred.

Getting started

Install the rfdetr package in a Python>=3.10 environment: bash pip install rfdetr To install from source (latest development version): bash pip install https://github.com/roboflow/rf-detr/archive/refs/heads/develop.zip

License and pricing

The open-source rfdetr package and Apache-designated models (RF-DETR-N, -S, -M, -L) are released under the Apache 2.0 license. Plus components (rfdetr_plus), including RF-DETR-Atto, -Femto, -Pico, -XL, and -2XL detection models, are licensed under PML 1.0. The license for the repository itself is Apache-2.0.

Based on the project's public README and repository metadata.