Detection answers is there a face here. Recognition answers the much harder question: whose face is it. That second question needs a comparison against a database of known identities — and that’s usually where it stops making sense to keep everything on-device, and starts making sense to lean on the cloud.
Why Recognition Moves to the Cloud
-
The reference database of known faces can grow into the thousands — too large to comfortably keep in memory on a small edge device
-
Recognition models are heavier than detection models, and benefit from real GPU inference
-
Centralizing recognition means the database of known identities can be updated in one place and instantly apply everywhere
-
Multiple devices (cameras at different doors, different sites) can share one consistent identity store
How the Pipeline Connects
-
The edge device detects a face locally (as covered in the previous post) and crops just that region — not the whole frame
-
That small crop is sent to a cloud recognition service, rather than a continuous video stream
-
The service computes a face embedding — a numeric vector that represents that face’s features
-
That embedding is compared against a database of known embeddings using a similarity search
-
A match (or «unknown») is returned to the device, typically in well under a second
What Actually Gets Sent
| On the Device | Sent to the Cloud |
|---|---|
| Full video stream | A single cropped face image, only when one is detected |
| Face detection (where) | Face recognition (who) |
| Runs continuously, every frame | Runs only on demand, per detected face |
Splitting the problem this way keeps bandwidth and cost down — the cloud only ever sees a cropped still image, not a live video feed — while still getting the accuracy and scalability of a proper recognition model and a centrally managed identity database. Detection stays fast and local; recognition stays accurate and centralized.
