← Thomas Yu

Ornimetrics

A bird feeder that knows who's eating. It names the species, feeds each one its own portion, spots birds that need help and learns its regulars, all on a Raspberry Pi.

Since
2024, founder and sole developer
Paper
First author · CS & IT Vol. 16 · ADCOM 2026
Patent
U.S. application filed 2026
Model
Open on Hugging Face

The feeder, modelled and animated in Spline. Drag to look around it.

The Ornimetrics feeder: black seed hopper on a camera housing over a square tray
The feeder from the side, camera window facing out
The feeder on a table by a window

Why

The end of the scramble

An ordinary tray is first come, first served: bullies crowd out timid birds, squirrels empty it, and leftover seed rots. Ornimetrics decides who eats.

Watchthe camera runs the detector continuously
Recognisebird, squirrel, person or dog, then the species
Feedthe disk turns out that species' portion
Logphoto, species and time, synced to the app
Cool downper species, then back to watching

In action

Watch it feed

The prototype turning out a portion, with the recognition feed on the laptop behind it.
The introduction video.

Parts tour

Four parts

Dispensing mechanism. A rotating disk with two pockets meters the seed out in exact amounts, and an incline under it pours the portion gently onto the tray. This is the mechanism behind the patent.

All four modelled and animated in Spline. Drag to turn them.

How it sees

One backbone, three answers

Cameralive feed at the feeder
DetectorYOLOv11n on the Hailo-8, about 30 fps
BackboneEfficientNetV2-S, each bird encoded once
Species555 North American species
Welfarespots birds that look hurt
Re-IDrecognises returning birds
The detector runs continuously on the Hailo-8; the three heads run on the Pi's CPU only when a bird is in frame, so the video never slows. A microphone path names birds by their calls.
0.93
mAP@50, the feeder detector
86.9%
top-1 over 555 species
0.963
AUC spotting injured birds, with no injured photos in training
45,000+
training images
ModelWhat it isResult
DetectorYOLOv11n, INT8, compiled for the Hailo-80.93 mAP@50 at about 30 fps
SpeciesEfficientNetV2-S fine-tuned on NABirds86.9% top-1 over 555 species
Species (open)The same network on Creative Commons photos302 species, licensed for commercial use
WelfareLearns what healthy birds look like and measures how far a new one sits from itAUC 0.963 injured, 0.974 dead
Re-IDArcFace embeddings with a gallery per speciesEnrols new regulars automatically

Published

The paper

Tested at the feeder itself: how often it named the bird, and how much seed each species left behind.

Correct identifications, 100 trials per species

Uneaten seed per feeding session, 15 trials

Sparrows waste the most, which is exactly what per-species portions are for.

CS & IT Conference Proceedings, Vol. 16 · presented at the 12th International Conference on Advanced Computing (ADCOM 2026), Vancouver, February 2026

Ornimetrics: Design and Evaluation of an AI-Enabled Smart Bird Feeder for Species Recognition, Waste Reduction, and Ecological Monitoring

Baichen Yu (first author), Tyler Boulom

What it grew into

Beyond the feeder

NextCOMU
→