roboflow/rf-detr

Built on a DINOv2 vision transformer backbone, the architecture unifies object detection and instance segmentation through a single API with support for fine-tuning across multiple model scales (N through 2XL). RF-DETR achieves 2-17ms inference latency on NVIDIA T4 with TensorRT optimization, with larger variants reaching 78.5 AP on COCO while maintaining real-time performance. The package integrates with Hugging Face and Roboflow's ecosystem, offering Apache 2.0 licensed base models alongside proprietary Plus variants under PML 1.0 licensing.

5,861 stars. Actively maintained with 106 commits in the last 30 days.