Every species. Every site. Every day.
Every bird that enters the frame gets a name, a timestamp, and the footage attached. The feed runs every day, dawn to dusk.
Probability a species is present in any given hour.
A camera at the feeder feeds 1080p video into a dedicated edge AI device. A real-time object detector runs on every frame, followed by a fine-grained species classifier trained on North American birds.
A visit is recorded when a bird is detected for at least three consecutive frames; the visit closes after five seconds without detection. A bird that leaves the frame and reappears within that window is counted as one continuous visit, not two.
Detections that fail confidence and frame-stability checks are flagged low-quality and excluded from the public counts. Each day's full broadcast is preserved on YouTube — the source footage for every observation.
This site monitors one elevated platform and tube feeder. The dataset is biased toward seed-eating passerines that visit feeders. Raptors, waterbirds, ground foragers, nocturnal species, and aerial insectivores are systematically underrepresented.
Origin · Beyond
My wife pulls up Birds of Virginia on her phone every time something new lands at our feeder. Nothing wrong with that. It just means that for the few seconds the bird is there, we are looking at the phone instead of at the bird.
So I built a camera that does the identifying, and a system around it that keeps doing it after we go inside. Every visit, every day, dawn to dusk. Species, time, duration, and the footage it came from, all of it written to a record we keep. Yesterday, at our feeder in Virginia, the camera logged 147 visits across six species. The first bird of the day was a Chipping Sparrow at 6:04 in the morning. House Finches spent the most time on camera, forty-one minutes in total. A White-breasted Nuthatch did not appear until 11:21 AM, and then came back twenty-nine times before evening.
Most of what is known about North American birds was learned from observations made by people. Millions of them, over more than a century, organized into the foundational datasets the field now runs on. What we're trying to add is the part those records can't easily contain. Continuous coverage of one place, on the same scale of time the birds themselves move at, with the footage attached to every claim.
One yard is the starting point. The shape we're working toward is one of these in every state — fifty fixed sites, continuously recording, all of them speaking the same data format, on the same clock. Whatever the long-tail uses turn out to be, they almost certainly require the network, not the node.
Or leave an email. We'll write when something happens.
Thanks. We'll be in touch.
Building something nearby? Want to host a node? anfon@birdsuntold.com