Salmon Detection
SalmonVision uses deep learning models to detect, track, and classify salmon in video feeds.
How Detection Works
- Frame Extraction — Video frames are captured at the configured FPS
- Object Detection — A YOLO-based model identifies salmon in each frame
- Tracking — A multi-object tracker assigns persistent IDs to each fish
- Counting — Fish are counted as they cross a virtual counting line
- Classification — Species classification runs on each detected fish
Detection Models
SalmonVision ships with pre-trained models for common scenarios:
| Model |
Use Case |
Speed |
Accuracy |
| SV-Nano |
Edge devices, real-time |
~60 FPS |
89% mAP |
| SV-Medium |
Balanced performance |
~30 FPS |
94% mAP |
| SV-Large |
Maximum accuracy |
~15 FPS |
97% mAP |
Supported Species
Out of the box, SalmonVision can classify:
- Chinook (King) Salmon
- Sockeye (Red) Salmon
- Coho (Silver) Salmon
- Pink (Humpy) Salmon
- Chum (Dog) Salmon
- Steelhead Trout
Adjusting Detection Settings
Navigate to Project Settings → Detection to configure:
- Confidence threshold — Minimum confidence to register a detection (default: 0.5)
- Counting line position — Drag to set the virtual line fish must cross
- Direction filter — Count upstream only, downstream only, or both
- Region of interest — Mask areas to ignore (e.g., riverbank, debris)
Review & Correction
Detections can be reviewed in the Review Queue:
- Open Dashboard → Review
- Thumbnail cards show each detection event
- Confirm, correct species, or reject false positives
- Corrections are used to improve model accuracy over time
Previous: Camera Setup | Next: Data Export