RT-DETR is the official implementation of the RT-DETR family of real-time object detectors — “DETRs Beat YOLOs on Real-time Object Detection” — showing that transformer-based detectors can match and surpass YOLO-style models at real-time speeds. The repository provides PyTorch and PaddlePaddle implementations, trained weights and training recipes, and the family has continued to grow with RT-DETRv2 and newer foundation-model-enhanced variants. It has about 5,600 stars on GitHub and remains actively maintained, with updates in 2026.
For years, real-time detection belonged to YOLO-style CNN detectors, while DETR transformers were accurate but slow and hard to train. RT-DETR closed that gap with an efficient hybrid encoder and IoU-aware query selection, delivering end-to-end detection — no non-maximum suppression post-processing — at real-time frame rates. The original paper appeared at CVPR 2024, and RT-DETRv2 improved the baseline with a collection of training refinements (“bag of freebies”). The repository also links the newest family members, including RT-DETRv4, which harnesses vision foundation models to boost accuracy without extra inference latency. Practical tooling is included, such as a utility for counting parameters and FLOPs, so the repo serves both as the research reference and a working starting point for production detection systems.
RT-DETR is free and open source under the Apache-2.0 license. Code and weights are available at no cost; you only provide the compute for training and inference.
RT-DETR is for computer-vision engineers and researchers building real-time detection for video analytics, robotics, autonomous systems and edge devices, especially those who want an end-to-end transformer detector instead of a YOLO pipeline.