# RT-DETR Training for OAK-D Camera Deployment

RT-DETR (Real-Time Detection Transformer) is Apache 2.0 licensed - **free for commercial use**. It's designed for real-time detection and works great on edge devices like the OAK-D 4 Pro.

## Why RT-DETR?

- ✅ **Apache 2.0 license** - truly free for commercial use
- ✅ **Excellent OAK camera compatibility** - exports cleanly to OpenVINO
- ✅ **Real-time performance** - 30-60 FPS on OAK-D 4 Pro
- ✅ **Modern transformer architecture** - competitive accuracy with YOLO
- ✅ **Easy deployment** - direct export to OpenVINO format

## Quick Start

### 1. Annotate Images

Use the annotation GUI:
```bash
.venv/bin/python annotation_gui.py
```

- Load your images from Settings
- Annotate knots manually or use auto-labeling
- Aim for 100+ annotated images for good results

### 2. Train Model

From the GUI:
1. Go to **Training** tab
2. Click "Prepare Dataset" (creates train/valid/test splits)
3. Select **RT-DETR** framework
4. Choose model size:
   - `nano` (r18): Fastest, 30-40 FPS on OAK
   - `small` (r34): Balanced
   - `medium` (r50): More accurate
   - `base` (l): Best accuracy, slower
5. Click "Start Training"

Or from command line:
```bash
.venv/bin/python train_rtdetr.py \
    --dataset-dir dataset_prepared \
    --model rtdetr-r18 \
    --epochs 100 \
    --batch-size 8
```

### 3. Test Model

```bash
.venv/bin/python predict_rtdetr.py \
    --weights runs/rtdetr_training/training/weights/best.pt \
    --image test_image.jpg
```

### 4. Export for OAK-D

Export to OpenVINO format:
```bash
.venv/bin/python export_rtdetr_oak.py \
    --weights runs/rtdetr_training/training/weights/best.pt \
    --img-size 640
```

This creates:
- `best_openvino_model/` - OpenVINO IR format (.xml + .bin files)
- `best.onnx` - ONNX format (intermediate)

### 5. Convert to Blob for OAK

**Option A: Online converter** (easiest)
1. Go to https://blobconverter.luxonis.com/
2. Upload `best_openvino_model/model.xml`
3. Select "OAK-D 4 Pro" 
4. Download `.blob` file

**Option B: Command line**
```bash
pip install blobconverter
blobconverter --openvino-xml best_openvino_model/model.xml \
              --shaves 6
```

### 6. Deploy to OAK-D Camera

Example DepthAI script:
```python
import depthai as dai
import cv2

# Create pipeline
pipeline = dai.Pipeline()

# Camera
cam = pipeline.createColorCamera()
cam.setPreviewSize(640, 640)
cam.setInterleaved(False)

# Neural network
nn = pipeline.createNeuralNetwork()
nn.setBlobPath("best.blob")
cam.preview.link(nn.input)

# Output
xout = pipeline.createXLinkOut()
xout.setStreamName("detections")
nn.out.link(xout.input)

# Run
with dai.Device(pipeline) as device:
    queue = device.getOutputQueue("detections")
    
    while True:
        detections = queue.get()
        # Process detections...
```

## Model Comparison

| Model | Size | Speed (OAK-D) | Accuracy | License |
|-------|------|---------------|----------|---------|
| RT-DETR r18 | ~15MB | 30-40 FPS | Good | Apache 2.0 ✅ |
| RT-DETR r34 | ~30MB | 20-30 FPS | Better | Apache 2.0 ✅ |
| YOLOv11n | ~6MB | 50-60 FPS | Good | AGPL ❌ |
| YOLOv6n | ~10MB | 40-50 FPS | Good | MIT ✅ |
| RF-DETR nano | ~15MB | 10-20 FPS* | Good | Check repo |

*May have compatibility issues with OpenVINO

## Training Tips

1. **Dataset size**: 
   - Minimum: 50 images
   - Good: 200+ images
   - Excellent: 1000+ images

2. **Data diversity**:
   - Different wood types
   - Various lighting conditions
   - Multiple knot sizes/types
   - Different angles

3. **Training settings**:
   - Start with `rtdetr-r18` for fastest iteration
   - Use `batch-size=8` if you have 8GB+ GPU
   - Train for 100-200 epochs
   - Use early stopping (patience=20)

4. **Data augmentation** (automatic):
   - Flips, rotations
   - Color adjustments
   - Crops and scales

## Troubleshooting

**Training is slow:**
- Reduce batch size
- Use smaller model (r18)
- Check GPU usage with `nvidia-smi`

**Low accuracy:**
- Add more training data
- Train longer (more epochs)
- Use larger model (r34 or r50)
- Check your annotations for errors

**OAK deployment fails:**
- Ensure OpenVINO export succeeded
- Check blob size (<200MB for OAK-D)
- Verify input size matches training (640x640)
- Try FP16 instead of FP32 to reduce size

## Resources

- [RT-DETR Paper](https://arxiv.org/abs/2304.08069)
- [Ultralytics RT-DETR Docs](https://docs.ultralytics.com/models/rtdetr/)
- [OAK-D Docs](https://docs.luxonis.com/)
- [DepthAI Examples](https://github.com/luxonis/depthai-experiments)

## License

RT-DETR is Apache 2.0 licensed - you can use it for:
- ✅ Personal projects
- ✅ Commercial products
- ✅ Internal business tools
- ✅ Proprietary software

No restrictions, no paid licenses required!