Pyyan / Compare / RF-DETR vs RTMDet vs Grounding DINO
RF-DETR vs RTMDet vs Grounding DINO
3 of 5Object Detection · verified 13 Aug 2026
| Specification | RF-DETR | RTMDet | Grounding DINO |
|---|---|---|---|
| Summary | First real-time detector past 60 mAP on COCO. | Pure throughput, and an MIT licence. | Detects whatever you describe in words. |
| COCO mAP | 60.5 | ~52 | ~52 zero-shot |
| Speed | Real time | 300+ | Moderate |
| Licence | Apache 2.0 | MIT | Apache 2.0 |
| Family | Transformer, set prediction | Single stage | Open vocabulary |
| Open vocabulary | No | No | Yes |
| Released | 2025 | 2022 | 2023 |
| Category | Object Detection | Object Detection | Object Detection |
| Official | Roboflow ↗ | OpenMMLab ↗ | IDEA Research ↗ |
Highlighted rows are where these differ.
RF-DETR
- Leads the RF100-VL domain-transfer benchmark
- DINOv2 backbone, markedly better on occluded objects
- Apache 2.0, which YOLO is not
Best for most custom detection tasks.
RTMDet
- Wins on raw speed where accuracy is sufficient
- MIT, so no derivative-work obligation
Best for high frame-rate video.
Grounding DINO
- Text prompt in, boxes out
- Fine-tuning still beats it on specialised objects
Best for classes you have no data for.