RF-DETR-N and RT-DETRv2-R18 land within 0.7 mAP of each other on COCO val2017: 51.4 vs 50.7 mAP@50-95 on an RTX 5070 Ti. Read that gap with the input sizes in mind: RF-DETR-N runs at 384 px and RT-DETRv2-R18 at 640 px, each at its author default. RT-DETRv2-R18 wins GPU speed, 51.1 vs 39.5 FPS, and leads on small objects. RF-DETR-N uses less VRAM and turns the speed result around on Raspberry Pi 5.
This is not a like-for-like resolution test. RF-DETR-N (30.47M parameters, Apache-2.0) is evaluated at 384 px; RT-DETRv2-R18 (20.18M, Apache-2.0) at 640 px, each at its default input size. Both run on COCO val2017 with verified rows on desktop GPU, Jetson Orin, and Raspberry Pi 5. Treat the accuracy numbers as each model at its own recommended setting, not as a controlled resolution sweep.
| Metric | RF-DETR-N | RT-DETRv2-R18 |
|---|---|---|
| mAP@50-95 | 5135.0 | 5075.0 |
| mAP@50 | 6990.0 | 6738.0 |
| mAP small | 2768.0 | 3348.0 |
| FPS (mean) | 39.5 | 51.1 |
| Total ms/image | 25.30 | 19.55 |
| Inference ms | 16.98 | 14.82 |
| Peak VRAM (MB) | 151 | 169 |
| Params (M) | 30.5 | 20.2 |
| GFLOPs | - | 60.0 |
| Input size | 384 | 640 |
| License | Apache-2.0 | Apache-2.0 |
Accuracy
mAP is shown in percent form. RF-DETR-N measures 51.4 mAP@50-95 to RT-DETRv2-R18's 50.7, a 0.7 point edge, but RT-DETRv2-R18 leads on small objects: 33.5 vs 27.7 mAP_small, a 17.32% relative gap. Its larger 640 px input is one plausible reason that row goes the other way. If small instances dominate your scenes, RT-DETRv2-R18 is the safer pick.
Speed
On the RTX 5070 Ti in PyTorch, RT-DETRv2-R18 runs 51.1 FPS to RF-DETR-N's 39.5, so it is 22.72% faster. The margin grows under conversion: 97.1 vs 65.7 FPS on ONNX Runtime and 125.5 vs 67.6 FPS on TensorRT FP32. On Jetson Orin the two are level at 10.3 vs 10.2 FPS.
The speed verdict flips on Raspberry Pi 5, where RF-DETR-N runs 2.1 FPS to RT-DETRv2-R18's 0.8. Its 384 px input helps most where compute is scarcest. If your target is a bare Pi, the desktop numbers point the wrong way.
- RF-DETR-N license
- Apache-2.0
- RT-DETRv2-R18 license
- Apache-2.0
- RF-DETR-N release
- 2026-02-23
- RT-DETRv2-R18 release
- 2024-05-01
- RF-DETR-N input
- 384 px
- RT-DETRv2-R18 input
- 640 px
- Evaluated weights
- LibreYOLO retrained checkpoints
When to pick which
Pick RF-DETR-N for CPU-bound edge targets like the Raspberry Pi 5, where its 384 px input makes it the faster detector, and where lower VRAM helps. Pick RT-DETRv2-R18 for GPU throughput and for small-object accuracy, where its 640 px input pays off. Both are Apache-2.0, so licensing does not force the choice.
Every number on this page comes from the verified dataset: same 500-image COCO val2017 slice, conf 0.001, IoU 0.6, max 300 detections, pycocotools mAP, identical protocol across all hardware and runtimes. The full protocol is on the methodology page. To rerun this comparison with your own filters, open compare. Accuracy is measured on LibreYOLO retrained checkpoints; other weight sources can yield different values.