Introduction

BirdWatchAI turns any camera pointed at your bird feeder into an automatic bird-watching station — identifying species on-device, saving photos and short video clips, and (optionally) sending you a notification when something rare shows up. This Part introduces the project, the two editions you can choose between, and the high-level picture of how a single detection happens.

What BirdWatchAI is

BirdWatchAI sits between an ordinary network camera and you. It watches the camera's video feed continuously, notices when something moves, decides whether the moving thing is actually a bird (filtering out squirrels, swaying branches, and shadows), identifies the species using a local AI model, and then saves a labeled snapshot plus a short clip — all without anything leaving your home network unless you opt in to community sharing.

Key things it does:

• Identifies species automatically — recognizes up to 965 North American birds with the default model, or 2,498 species worldwide with the optional larger model.
• Captures the moment — high-resolution snapshot and a short video clip with the species, confidence, timestamp, temperature, and location burned in as an overlay.
• Tells you what's rare — compares each species against regional / seasonal data so a Wood Thrush stands out from a House Sparrow.
• Notifies you — email, mobile push (Pushover or ntfy), Windows toast, or even a digital photo frame.
• Builds a journal — a searchable detection history, statistics dashboards, a life list, and (in the desktop app) gamified achievement badges.
• Connects you to others — optionally publish sightings to the worldwide community feed at www.birdwatchai.com.

Desktop vs. Server — which one should I run?

BirdWatchAI ships in two flavors. They share the same AI models and the same community backend; they differ mainly in where and how they run.

You can run either — never both pointing at the same camera at the same time, since most cameras only allow a small handful of simultaneous RTSP streams. Pick the one that fits your situation; the rest of this manual flags instructions that apply to only one edition.

Feature matrix

What's available where:

System requirements

Windows desktop app

Server (Raspberry Pi or Docker)

How a detection happens

Both editions run the same pipeline. When monitoring is active:

1. Motion is detected — either by the camera's own hardware (via ONVIF, recommended for Tapo) or by the app comparing low-resolution frames.
2. A high-resolution snapshot is captured immediately, and a short video clip starts recording (using a rolling pre-buffer so the clip includes the seconds before the trigger).
3. Best-frame extraction picks the sharpest frame from the clip — important when the bird arrives a moment after the motion trigger.
4. An object detector finds the bird, crops to it, and silently drops non-bird motion like squirrels or branches.
5. The classifier runs the local AI model on the cropped bird and returns the most likely species plus a confidence percentage.
6. Enrichment looks up the current outdoor temperature and the regional rarity of the species (when your location is set).
7. If confidence is above the threshold, the snapshot is saved with an on-image overlay, the clip is labeled, the detection is added to history, and any enabled notifications fire. Below-threshold detections are discarded by default (or kept in a "for review" folder if you opt in).

That whole pipeline runs in a couple of seconds per detection. The rest of this manual is about getting the camera connected, tuning the thresholds, and reading the results.