Pigeons, Rats, and Python
Computer Vision for the Concrete Jungle
Maria Ashby · Founder, 3percentclub · July 10, 2026
What is Computer Vision?
Teaching machines to see the city the way an ecologist would.
Classification
Names what is in a picture. "This is a bird."
Object Detection
Finds where objects are and boxes them.
Segmentation
Traces an object's exact shape, pixel by pixel.
What is a YOLO model?
"You Only Look Once" reads a whole frame in a single glance: ultra-fast, lightweight, and runs live on low-cost devices.
Case Study: Billion Oyster Project
The Problem
Manual wildlife counts don't scale. Nobody can sit underwater counting oysters 24/7.
The AI Solution
Fine-tuned an open-source model (YOLOv8) on real New York Harbor footage.
The Takeaway
If it works in murky water, it works on your block.
85%
Accuracy in murky harbor water
Inside the Live Tracker
How the pipeline works, simply put.
Programmatic Sourcing
Pulls live, open ecological data automatically.
Semantic Costumes
Relabels off-the-shelf data for local species without custom training.
Privacy by Design
Blurs human faces right at the start to protect neighbors.
Auditing the AI Footprint
Balancing conservation with compute cost.
Average Energy / Prompt
0.3 Wh
0.0003 kWh
What is the average energy use per AI prompt?
Average Water / Prompt
0.5 ml
evaporated server cooling water
What is the average water use per AI prompt?
The Comparison
330 prompts
= 100 Wh of grid power
How do they compare to watching television? 330 prompts uses the same electricity as 1 hour of streaming on a smart TV.
Build it on your block
Everything here is open. Grab the code, run the numbers, or book a call to bring it to your community.