roboflow/rf-detr
RF-DETR is a real-time object detection and segmentation model architecture developed by Roboflow, SOTA on COCO, designed for fine-tuning. [ICLR 2026]
What it solves
RF-DETR addresses the need for high-accuracy, real-time computer vision models that can handle multiple tasks—object detection, instance segmentation, and keypoint detection—without sacrificing latency. It provides a scalable set of model sizes (from Nano to 2XLarge) to allow developers to balance precision and speed based on their hardware constraints.
How it works
RF-DETR is a real-time transformer architecture built on a DINOv2 vision transformer backbone. It uses a consistent API for its various tasks and was developed using neural architecture search (NAS) to optimize the model sizes for performance. The project provides an open-source package (rfdetr) and supports integration with the Inference library for streamlined deployment.
Who it’s for
Computer vision engineers and developers who need state-of-the-art (SOTA) real-time detection and segmentation models for deployment on edge devices or high-performance servers.
Highlights
- Multi-task Support: A single API for object detection, instance segmentation, and keypoint detection.
- DINOv2 Backbone: Leverages a powerful vision transformer backbone for improved accuracy.
- Optimized Performance: Achieves competitive accuracy-latency trade-offs compared to YOLO and other DETR-based models.
- NAS-Driven: Model architectures were created using neural architecture search to ensure optimal efficiency.