RT-DETR
The first real-time end-to-end detection transformer: DETR accuracy at YOLO speeds, with no NMS.
About
RT-DETR is Baidu's answer to the complaint that transformers are too slow for real-time detection. It keeps the end-to-end DETR design, no NMS post-processing and no anchor hand-tuning, but adds an efficient hybrid encoder that processes multi-scale features with decoupled intra-scale interaction and cross-scale fusion, plus uncertainty-minimal query selection for the decoder. The result is a transformer detector that genuinely runs in real time.
The authors report RT-DETR-R50 hitting 53.1% AP on COCO val2017 at 108 FPS on a T4 GPU with TensorRT FP16, and the HGNetv2-X variant reaching 54.8% AP at 74 FPS, outperforming same-scale YOLO detectors on both speed and accuracy at release. It is Apache-2.0 licensed, ships in Paddle and PyTorch, and Hugging Face carries Transformers-native checkpoints (PekingU/rtdetr_r50vd), which made it the go-to real-time transformer detector.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported: RT-DETR-R50 53.1 AP at 108 FPS; RT-DETR-HGNetv2-X 54.8 AP at 74 FPS on COCO val2017 (T4, TensorRT FP16)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: paper (arXiv:2304.08069, CVPR 2024) with the full COCO/FPS table confirmed via the official repo README; Apache-2.0 confirmed via Hugging Face model cards and repo license badge. Strengths: landmark real-time DETR, permissive license, Transformers-native checkpoints, Ultralytics support. Docked for: (1) original repo is one maintainer's project (lyuwenyu), less institutional backing than Ultralytics; (2) newer YOLO generations have since closed the gap on some size classes. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · CVPR 2024 (arXiv 2304.08069, Apr 2023)
- · R50 53.1 AP at 108 FPS / HGNetv2-X 54.8 AP at 74 FPS on COCO val2017 (T4, TensorRT FP16), author-reported
- · Transformers-native checkpoints PekingU/rtdetr_r50vd