# RF-DETR for TI EdgeAI
### Real-Time Transformer Detection with a DINOv2 Backbone
[](https://opensource.org/licenses/Apache-2.0)
[](https://onnx.ai/)
[](https://github.com/TexasInstruments/edgeai)
[](https://cocodataset.org/)
---
## Overview
**RF-DETR** is a real-time transformer-based object detection (and instance segmentation) architecture developed by Roboflow, achieving state-of-the-art accuracy/latency trade-offs on COCO (presented at ICLR 2026). It builds on a **DINOv2 ViT backbone** paired with a DETR-style decoder and a hierarchical feature pyramid, giving it strong small-object and dense-scene performance without the NMS and anchor-tuning overhead of traditional detectors.
RF-DETR ships in six size variants, Nano through 2XLarge, for flexible accuracy-speed trade-offs. The Nano through Large variants are released under Apache 2.0; XLarge and 2XLarge require the `rfdetr[plus]` extra and are licensed under PML 1.0. This folder packages the ONNX exports and TIDL configs for the Apache-licensed detection variants (Nano/Small/Medium/Large); the `prepare_model.py` script can additionally export the PML-licensed XLarge/2XLarge detection variants and the segmentation family on request.
---
## Model Variants
| Model | Input Size | Reference mAP[.5:.95]% | Reference mAP[.50]% | Validated Devices | Config |
|-------|-----------|--------------|-----------|--------------------|--------|
| `rfdetr_nano` | 384×384 | 48.4 | 67.6 | TDA4VH | [rfdetr_nano_config.yaml](rfdetr_nano_config.yaml) |
| `rfdetr_small` | 512×512 | 53.0 | 72.1 | TDA4VH | [rfdetr_small_config.yaml](rfdetr_small_config.yaml) |
| `rfdetr_medium` | 576×576 | 54.7 | 73.6 | TDA4VH | [rfdetr_medium_config.yaml](rfdetr_medium_config.yaml) |
| `rfdetr_large` | 704×704 | 56.5 | 75.1 | TDA4VH | [rfdetr_large_config.yaml](rfdetr_large_config.yaml) |
> mAP values are on COCO val2017. `rfdetr_xlarge` (700×700, mAP[.5:.95] 58.6) and `rfdetr_2xlarge` (880×880, mAP[.5:.95] 60.1) are available via `prepare_model.py --plus` but are licensed under PML 1.0 and are not shipped as ONNX/config files in this folder.
**Recommended for edge deployment:** `rfdetr_nano` (best accuracy/compute trade-off, smallest)
---
## Quick Start
### Prerequisites
```bash
# ONNX export dependencies
pip install "rfdetr[onnx]"
# For XLarge / 2XLarge variants (PML 1.0 license)
pip install "rfdetr[onnx,plus]"
# Inference and deployment
pip install onnx>=1.22.0
pip install onnxruntime>=1.23.2
```
### Export the Model
```bash
# List all available variants with accuracy and latency info
python prepare_model.py --list-models
# Export the default model (rfdetr_nano)
python prepare_model.py
# Export a specific model variant
python prepare_model.py --model rfdetr_medium
# Export multiple variants at once
python prepare_model.py --model rfdetr_nano rfdetr_small rfdetr_medium rfdetr_large
# Export XLarge / 2XLarge (requires rfdetr[plus], PML 1.0 license)
python prepare_model.py --model rfdetr_xlarge rfdetr_2xlarge --plus
```
The script automatically:
- Installs `rfdetr[onnx]` (or `rfdetr[onnx,plus]` for XLarge/2XLarge) if not already present
- Downloads pretrained COCO weights from HuggingFace on first use
- Exports the selected variant(s) to ONNX (opset 17 by default, static batch dimension)
- Saves the result as `rfdetr_.onnx` in the output directory
### Compile and Infer uing edgeai-tidlrunner
> **Note:** Run the commands below from inside the `tidlrunner` directory (the cloned [edgeai-tidlrunner](https://github.com/TexasInstruments/edgeai-tidlrunner) repository), with `--config_path` pointing to this model's config file.
**Compile using edgeai-tidlrunner - on PC**
```bash
cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device J784S4 \
--config_path /path/to/rfdetr_nano_config.yaml
```
**Run Inference Benchmark - on device**
```bash
cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device J784S4 \
--config_path /path/to/rfdetr_nano_config.yaml
```
### Compile and Infer using edgeai-tidl-tools (Advanced):
Follow the instructions at https://github.com/TexasInstruments/edgeai-tidl-tools
### Deploy using edgeai-tidl-tools:
Deplyment can be done using **[edgeai-tidl-tools](https://github.com/TexasInstruments/edgeai-tidl-tools)**. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.
---
## Citation
If you use RF-DETR in your research, please cite:
```bibtex
@software{rfdetr2025,
title = {RF-DETR},
author = {Robinson, Isaac and Robicheaux, Peter and Popov, Matvei},
year = {2025},
publisher = {Roboflow},
url = {https://github.com/roboflow/rf-detr},
note = {International Conference on Learning Representations (ICLR) 2026}
}
```
---
## 🔗 Resources
| Resource | Link |
|----------|------|
| **RF-DETR Source Code** | [roboflow/rf-detr](https://github.com/roboflow/rf-detr) |
| **RF-DETR Documentation** | [rfdetr.roboflow.com](https://rfdetr.roboflow.com) |
| **RF-DETR on HuggingFace** | [huggingface.co/roboflow](https://huggingface.co/roboflow) |
| **edgeai-tidl-tools** | [GitHub](https://github.com/TexasInstruments/edgeai-tidl-tools) |
| **edgeai-tidlrunner** | [GitHub](https://github.com/TexasInstruments/edgeai-tidlrunner) |
| **EdgeAI SDK** | [Documentation](https://github.com/TexasInstruments/edgeai/blob/main/edgeai-mpu/readme_sdk.md) |
---
## Related Models