RT-DETR

RT-DETR is the first real-time end-to-end object detector to outperform YOLO architectures by eliminating non-maximum suppression (NMS) via vision transformers.

Developed by Baidu, RT-DETR (Real-Time DEtection TRansformer) solves the latency bottleneck of traditional detectors by using a hybrid encoder and IoU-aware query selection. It achieves 54.8% AP on COCO val2017 while running at 114 FPS on an NVIDIA A100, effectively replacing hand-crafted NMS post-processing with a streamlined transformer pipeline. This architecture allows for flexible speed-accuracy trade-offs without retraining (simply by adjusting decoder layers) making it the benchmark choice for high-performance computer vision deployment.

https://github.com/lyuwenyu/RT-DETR