🦴 YOLO11 — Fracture Detection (Freeze-10, FP16 ONNX)
This model fine-tunes YOLO11n for fracture detection using the
Fracture Dataset (Roboflow).
The first 10 layers were frozen to retain pretrained detection features, and the model was exported to ONNX (FP16) for deployment.
⚙️ Configuration
| Attribute | Value |
|---|---|
| Base Model | yolo11n.pt |
| Dataset | Fracture (Roboflow) |
| Epochs | 30 |
| Batch Size | 32 |
| Image Size | 640×640 |
| Optimizer | Auto |
| Freeze Layers | 10 |
| Precision | FP16 (half=True) |
| Export Format | ONNX |
| Device | GPU (0,1) |
🩺 Example Detection
📈 Results
| Metric | Value |
|---|---|
| mAP50 | 0.920 |
| mAP50-95 | 0.524 |
| Precision (B) | 0.903 |
| Recall (B) | 0.832 |
| Inference Time (ms) | 33.04 |
| FPS | 30.27 |
| Model Size (MB) | 5.2 |
FP16 inference maintained identical accuracy to FP32 while reducing latency.
Layer freezing improved training stability and avoided overfitting on the limited dataset.