Custom Models

SalmonVision supports custom model training for scenarios not covered by the default models.

When to Train a Custom Model

Consider custom training when:

• Your camera setup has unusual lighting or angles
• You need to detect additional species or life stages (e.g., smolts, kelts)
• You want to improve accuracy for a specific river system
• You need to detect non-salmon species (e.g., lamprey, trout)

Preparing Training Data

Using SalmonVision Annotations

1. Go to Dashboard → Review and review detection results
2. Confirmed and corrected detections are automatically added to your training dataset
3. Export your annotations via Settings → Training Data → Export COCO

Manual Annotation

You can also prepare data externally:

1. Export video frames from your cameras
2. Annotate using tools like CVAT or Label Studio
3. Export in COCO format
4. Upload via Settings → Training Data → Import

Training a Model

1. Navigate to Settings → Custom Models
2. Click New Training Job
3. Configure training parameters:

base_model: sv-medium
epochs: 100
batch_size: 16
learning_rate: 0.001
augmentation: true
validation_split: 0.2

4. Click Start Training

Training runs on SalmonVision’s cloud GPU infrastructure. You’ll receive an email notification when training completes.

Deploying a Custom Model

1. Go to Settings → Custom Models
2. Find your trained model and click Deploy
3. Select which project(s) and camera(s) should use the model
4. Click Activate

Model Performance

After deployment, track your model’s performance in the Model Metrics panel:

• Precision & Recall curves
• mAP scores over time
• Confusion matrix by species
• Comparison against the base model

Previous: API Reference | Next: FAQ