What it does BerryNet turns a Raspberry Pi (or similar edge device) into a local deep-learning gateway. It ingests camera streams or image files, runs inference via MQTT-connected components, and spits results to a web dashboard. The default stack bundles a camera client, an object-detection engine (TinyYOLO or MobileNet SSD), and a Freeboard dashboard. Everything stays on-device; no internet connection is needed.

The interesting bit The project treats the Pi as a first-class citizen, not a toy. It wires together Node.js/Python services over MQTT so you can swap engines or add collectors without rewriting the pipeline. Multiple devices can form a distributed AIoT network—useful if you want a cluster of cheap sentinels instead of one overworked board.

Key highlights

• Supports local USB, Pi, and IP cameras with adjustable frame rates
• Default engines: Inception v3 for classification, TinyYOLO or MobileNet SSD for detection
• Includes data collector to save snapshots and inference results to disk
• Components managed via supervisor; configs live in /etc/supervisor/conf.d/
• Provides retraining path for custom YOLO models via darknet

Caveats

• README still references TinyYOLO and Inception v3; newer model support is unclear
• Dashboard setup requires manual JSON editing for remote viewing
• “Email and IM notification” mentioned as future work with no visible progress

Verdict Good fit for hobbyists or small shops building offline surveillance or quality-check stations on a budget. Skip it if you need managed cloud scaling or modern transformer-based models.