RT-DETRv2 R50vd
PekingU/rtdetr_v2_r50vd
published Jan 2025 · updated Feb 2025
RT-DETRv2 R50vd is a real-time object detection transformer that improves accuracy and deployment flexibility using selective multi-scale feature extraction and a discrete sampling operator.
specs
| Task | Object Detection |
| Architecture | RT-DETRv2 with ResNet-50vd backbone |
| License | Apache 2.0 |
about this model
RT-DETRv2 R50vd is a real-time object detection transformer model hosted on Gigarouter's managed API. It refines the RT-DETR architecture by introducing selective multi-scale feature extraction, an optional discrete sampling operator that removes deployment constraints specific to DETRs, and improved training strategies including dynamic data augmentation and scale-adaptive hyperparameters. These enhancements increase flexibility and practicality while preserving real-time inference speed.
Performance
Trained on COCO train2017 and evaluated on COCO val2017, RT-DETRv2 consistently outperforms its predecessor across all model sizes at equal speeds. The small variant (RT-DETRv2-S) achieves 48.1 mAP, a +1.6 improvement over RT-DETR-R18.
Key Strengths
- Selective multi-scale feature extraction via distinct sampling points per feature scale in deformable attention.
- Discrete sampling operator (optional) replacing
grid_sampleto broaden deployment compatibility. - Dynamic data augmentation and scale-adaptive hyperparameters improve accuracy without latency cost.
The model is licensed under Apache 2.0 and is available for inference via Gigarouter's OpenAI-compatible API endpoint.
best for
- ·Autonomous driving
- ·Surveillance systems
- ·Retail analytics
FAQ
It is ideal for real-time object detection in autonomous driving, surveillance, robotics, and retail analytics.
It is released under the Apache 2.0 license.
It takes images as input and returns bounding boxes with class labels and confidence scores.
Use the gigarouter OpenAI-compatible endpoint with your API key.
It outperforms RT-DETR across all sizes while maintaining the same real-time speed, thanks to improved training strategies and deployment-friendly design.
We're benchmarking and onboarding RT-DETRv2 R50vd as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.