Technology
The goal isn't to replace field work. It's to make field work more informed, more efficient, and more scalable.
Project Snapshot
Training Dataset
7,500+
Annotated underwater images
Recognition Classes
9 fish classes
Including life-stage variants
Current Focus
Model performance
Expanded datasets & field validation
Edge Computing
Early R&D
Research and development underway
Field Testing
Underway
Independent data collection ongoing
Philosophy
Learn, iterate,
improve
Built in the open
Fish Recognition
Current Sight's first computer vision model began with a single objective: identify invasive Sacramento pikeminnow in underwater video. As the project evolved, so did the model.
Today, it recognizes nine visual classes of fish — including separate life stages when appearance changes significantly, such as juvenile and adult salmonids. Rather than forcing every fish into a single category, the model is being trained to recognize the full range of what conservation professionals actually see in the field.
Every new training cycle is an opportunity to improve. More representative data. Better annotations. Better questions. Better results.
Recognition Classes
9 TotalData Matters
One of the biggest lessons learned during development has been that the quality of the model is inseparable from the quality of the data behind it.
Building the dataset has required much more than collecting images. It has meant learning how fish behave in different habitats, understanding how lighting, water clarity, camera position, and fish orientation affect model performance, and organizing thousands of images so every annotation can be confidently verified.
Every image teaches something.
Built for the Field
Technology developed in an office isn't enough. Current Sight is being built with real-world field deployment as the end goal — not just a capable model, but a system that works where it needs to work.
Every field season raises new questions. Those questions shape the next generation of the technology.
Computer vision model identifies fish species from underwater video footage.
Diverse training datasets and iterative retraining cycles improve accuracy over time.
Research underway for remote, low-power deployment in field monitoring environments.
Adding context beyond imagery — water temperature, flow, and conditions data.
Outputs designed to support real fisheries monitoring and restoration reporting needs.
Measuring Progress
Current Sight evaluates model performance using industry-standard computer vision metrics. These aren't just numbers to report — they guide every future data collection and model development decision.
The objective isn't simply to improve numbers. The objective is to build a tool conservation professionals can trust.
| Metric | Why It Matters |
|---|---|
| Precision | Measures how often identified fish are classified correctly. |
| Recall | Measures how consistently the model detects fish that are actually present. |
| F1 Score | Balances precision and recall into a single performance measure. |
| mAP | Provides an overall measure of detection and classification performance. |
Project Timeline
Current Sight is intentionally building incrementally — proving each capability before moving to the next. Every milestone answers one question while creating the next.
That's exactly how meaningful innovation happens.
2026 milestone
2026 milestone
2026 milestone
2026 milestone
2026 milestone
Next up
Ahead
Ahead
The mission won't. If you're working in fisheries conservation and curious about what practical computer vision could do for your work, we'd love to talk.
Start a conversation →