Bird Feeder Detector.

What it does

1. Capture. A Raspberry Pi 2 with a Logitech webcam watches the feeder. Every few seconds it compares the new frame to the last one, and saves it only if enough pixels changed: a bird arrived, moved or left.
2. Collect. New frames go to the cloud, first through Google Drive with rclone, later straight into Roboflow with its SDK.
3. Label. In Roboflow, auto-label draws rough boxes and I correct them by hand.
4. Train. The labeled set goes to a Google Colab notebook, which fine-tunes YOLOv8n on a free T4 GPU.
5. Detect. On my MacBook, inference.py runs the trained model on the webcam feed through PyTorch’s Apple GPU backend. It draws a box with the species and confidence on each bird, and logs every session to Weights & Biases.

How it works

The work is split by what each machine is good at. The Pi 2 is far too slow to run a neural network, so it only captures. That also makes it simple enough to leave running unattended for days. Training happens on a borrowed cloud GPU, and live detection runs on the laptop.

Some choices along the way:

• Object detection, not classification. Frames often hold more than one bird, so I needed a location and species for each bird, not one label per photo.
• Colab over a home GPU. I planned a setup on a GTX 1050 Ti at my parents’ house, with Docker, CUDA, a Tailscale VPN and rclone. That would have been a more impressive infrastructure story, but Colab’s free T4 was faster and needs zero maintenance.
• Roboflow’s free tier. It made labeling much faster. The trade-off is that the dataset is public on Roboflow Universe, since the private plan is $249 a month.

Results

Both models are YOLOv8n trained for 50 epochs at 640 px. These are each model’s scores on its own validation set:

Validation scores only tell part of the story, though. The live tests below told me more.

What broke