Pyyan / Compare / OWLv2 vs RF-DETR vs RTMDet
OWLv2 vs RF-DETR vs RTMDet
3 of 5Object Detection · verified 13 Aug 2026
| Specification | OWLv2 | RF-DETR | RTMDet |
|---|---|---|---|
| Summary | Open-vocabulary detection trained with self-labelling. | First real-time detector past 60 mAP on COCO. | Pure throughput, and an MIT licence. |
| COCO mAP | ~45 zero-shot | 60.5 | ~52 |
| Speed | Moderate | Real time | 300+ |
| Licence | Apache 2.0 | Apache 2.0 | MIT |
| Family | Open vocabulary | Transformer, set prediction | Single stage |
| Open vocabulary | Yes | No | No |
| Released | 2023 | 2025 | 2022 |
| Category | Object Detection | Object Detection | Object Detection |
| Official | Google ↗ | Roboflow ↗ | OpenMMLab ↗ |
Highlighted rows are where these differ.
OWLv2
- Accepts an image as the query, not just text
Best for querying by example image.
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.