Hummingbird Classifier
A motion-triggered camera and a fine-tuned image model that identifies which species is at the feeder, running entirely on a Raspberry Pi in the garden.
A camera pointed at a feeder, a small convolutional model, and a nightly job that sorts the day's captures by species.
Why
I wanted to know how many individual birds were visiting, and counting by eye through a window turned out to be hopeless — they are fast, they look alike, and they do not hold still.
How it works
The Pi captures on motion, discards anything without a bird in it, and pushes the survivors to a nightly classification batch. Results land in Home Assistant as sensor values, which means the garden now has a dashboard.
What was hard
Negative examples. A model trained only on birds finds birds everywhere, including in an empty frame at dusk. Roughly two thirds of the final training set is deliberately nothing happening.
Built with
- Python
- PyTorch
- Raspberry Pi
- Home Assistant