// the find
DT42/BerryNet
Deep learning gateway on Raspberry Pi and other edge devices
BerryNet turns a Raspberry Pi (or similar edge device) into a local inference gateway: camera in, MQTT bus in the middle, inference engine out, results on a web dashboard. It's aimed at hobbyists and prototypers who want object detection or classification running fully offline on cheap hardware. Last commit was February 2023, so treat it as a reference architecture, not an actively maintained library.
The MQTT-bus design is genuinely good — every component (camera client, inference engine, dashboard, notifier) is decoupled and swappable without touching the others. Multi-backend engine support (TFLite, OpenVINO, Darknet, Caffe, Movidius) is real, not superficial — each has its own service module. Supervisor-based process management is the right call for embedded Linux, no container overhead needed. The data collector that saves snapshots + inference results to disk is a practical detail most similar projects omit.
Dead in the water since early 2023 — TFLite 2.x API has breaking changes, Movidius/NCS support is long EOL, and the pinned model weights link to a Google Drive folder that may or may not still exist. The mixed Node.js/Python split (berrynetjs/ vs berrynet/) creates two separate dependency trees and two runtimes for what is essentially glue code — there's no obvious reason both exist at this stage. No authentication on the MQTT broker or the Freeboard dashboard by default, which matters the moment the device is on any shared network. Setup is a shell script with a `./configure` that pulls in system packages non-deterministically; Docker support exists but is an afterthought, not the primary path.