Models & Weights

The system is designed to be model-agnostic within the YOLO family.

Supported Architectures

• YOLOv8: The previous standard. Good balance.
• YOLOv11 (Default): State-of-the-art accuracy/speed ratio.
• YOLO-NAS: Supported via specific export adapters.

Updating the Model

To update the AI brain without code changes:

1. Train/Export your model to a .pt file (PyTorch).
2. Place the file in services/camera-manager (mapped volume).
3. Update .env:

   MODEL_PATH=/app/new-model-v2.pt

4. Restart the batch-inference container.
   • Note: The first startup will take longer as it recompiles the TensorRT engine for the new weights.

Classes

The default model is trained on the FirstBreath Dataset with the following class IDs:

ID	Class Name	Description
0	Horse	The main subject.
1	Person	Used to pause alerts (maintenance mode) when a human is present.