rfdetr

What it is and what it does

RF-DETR is a production-ready computer vision library for detecting objects, segmenting instances, and detecting keypoints in images using a transformer architecture. It wraps a DINOv2 vision transformer backbone and provides a unified API across all three tasks. The library is built by Roboflow and includes benchmarked model sizes (N, S, M, L) under Apache 2.0, plus larger variants (XL, 2XL) under PML 1.0.

The package is designed for developers building real-time vision applications who need strong accuracy-latency tradeoffs. It depends on PyTorch, torchvision, transformers, numpy, and several utility libraries (requests, tqdm, pydantic, supervision, pyDeprecate). The library is actively maintained, recently released, and has no known security vulnerabilities.

Use it for:

• Build real-time object detection systems for robotics or autonomous systems using the lightweight N/S/M model variants.
• Perform instance segmentation on images or video streams where you need both bounding boxes and pixel-level masks.
• Integrate keypoint detection (preview) into pose estimation or human activity recognition pipelines.
• Benchmark detection accuracy against other SOTA models using the published COCO and RF100-VL results.
• Deploy models with neural architecture search to optimize for your specific dataset and hardware constraints.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

RF-DETR is a transformer-based object detection, instance segmentation, and keypoint detection library built on a DINOv2 vision transformer backbone, designed for real-time inference.

Yes. RF-DETR is actively maintained, recently released, has no security vulnerabilities, and offers strong accuracy-latency tradeoffs for object detection and segmentation. The Apache 2.0 base license is permissive. Install friction is low. The main consideration is Python 3.10+ requirement and PyTorch dependency; if your environment already supports those, this is a solid choice for real-time vision tasks.

Install

rfdetr on PyPI

pip install rfdetr

uv add rfdetr

poetry add rfdetr

Installing rfdetr

Before you install

Installation is straightforward via pip with low friction. The package is actively maintained with a recent release (3 days old) and strong repository activity (8972 stars). It requires Python 3.10 or later and depends on established libraries like torch, torchvision, transformers, and numpy.

License in practice

The base rfdetr package and standard models are Apache 2.0 licensed, permitting commercial and private use. Premium models (RF-DETR-XL/2XL) use PML 1.0, which may have different terms—verify licensing for those components if you plan to use them.

Quickstart

pip install rfdetr

import rfdetr
model = rfdetr.load_model('rf-detr-s')
results = model.predict(image_path='photo.jpg')

Requires Python 3.10 or later. Torch and torchvision must be installed; GPU support is optional but recommended for real-time performance.

Verify before relying

• Whether keypoint detection is production-ready or still in preview as stated in the description
• Specific hardware requirements or recommended GPU types for real-time inference targets
• Whether the package includes pre-trained model weights or if they must be downloaded separately

Package facts

Evidence: rfdetr-1.9.2-py3-none-any.whl

Keywords: machine-learning, deep-learning, vision, ML, DL, AI, DETR, RF-DETR, Roboflow

Tags

More Software Development packages