FinShield

Inspiration

Rob Stewart's film Sharkwater and the fact that 73 million sharks die each year due to the demand for their fins.

What it does

-Uses YOLOv8 to spot finning or fins in video or satellite images -Cross references AIS vessel patterns (loitering, drops) -Tags suspicious vessels and alerts nearby authorities when our model has high enough confidence

How we built it

-Annotated 1800+ finning frames in Roboflow -Trained YOLOv8 CV model in Colab -Pulled AIS data via Global Fishing Watch API
-Integrating the model and API in the backend of our react app

Challenges we ran into

-Few clear finning images -Waves and boat shadows -high false positives, so more training -Deploying website on github -Debugging code

Accomplishments that we're proud of

-Creating open-source software on GitHub -Over 80% detection rate on some videos -Integrating API's and CV models on the backend that work accurately -Building a quality product in our first hackathon

What we learned

-AIS context cuts false alarms
-How to tweak parameters when training, and reinforcing the model for better accuracy -Real world edge cases need ongoing retraining -How to deploy apps -Using git commands for version control

What's next for Finshield

-Apply to ocean-tech accelerators
-Partner with NGOs and enforcement teams -Train model for accuracy even with obscurity -Adding features like user reporting to the dashboard -Scale operations to worldwide ocean monitoring