| Property | Value |
|---|---|
| Parameter Count | 43M parameters |
| License | Apache-2.0 |
| Paper | DETRs Beat YOLOs on Real-time Object Detection |
| Performance | 53.1% AP on COCO, 108 FPS on T4 GPU |
What is rtdetr_r50vd?
RT-DETR (Real-Time Detection Transformer) is a groundbreaking object detection model that bridges the gap between DETR's accuracy and YOLO's speed. Developed by researchers at Peking University, it's the first real-time end-to-end object detector that eliminates the need for Non-Maximum Suppression (NMS) while maintaining high performance.
Implementation Details
The model utilizes a hybrid architecture combining an efficient hybrid encoder with uncertainty-minimal query selection. It processes multi-scale features through two key components: Attention-based Intra-scale Feature Interaction (AIFI) and CNN-based Cross-scale Feature Fusion (CCFF). Images are preprocessed to 640x640 pixels with specific normalization parameters.
- Trained on COCO 2017 dataset (118k training images)
- Supports flexible speed tuning through adjustable decoder layers
- Achieves 53.1% AP on COCO validation set
- Operates at 108 FPS on T4 GPU
Core Capabilities
- Real-time object detection with state-of-the-art accuracy
- End-to-end detection without NMS post-processing
- Multi-scale feature processing
- Flexible speed-accuracy trade-off
Frequently Asked Questions
Q: What makes this model unique?
RT-DETR uniquely combines transformer-based detection with real-time performance, outperforming both YOLO models in speed and accuracy while eliminating the need for NMS. It's 21 times faster than DINO-R50 while achieving better accuracy.
Q: What are the recommended use cases?
The model is ideal for real-time object detection applications requiring both speed and accuracy, such as surveillance systems, autonomous driving, and real-time video analysis. Its flexible architecture allows for deployment in various scenarios with different speed-accuracy requirements.