Ayonix vs cloud-only face recognition

Which requirement settles the choice?

Three requirements decide the architecture on their own, and none of them is a performance question. If any one of them applies, the shortlist is already narrowed before a single product is evaluated.

• A legal or policy rule that images or templates may not leave the premises. Common in defence, central banking, and any deployment covered by a data residency clause.
• A requirement that recognition continues during an internet outage, which a cloud API cannot meet because the matching happens elsewhere.
• A latency budget below the round trip to the nearest vendor region, typically anything that has to open a gate at walking pace.

Where none of the three applies, the decision is genuinely open and the arithmetic below decides it. Where one does, the remaining question is which on-premise system to choose, not whether to have one.

How do the two architectures compare?

Nothing in that table is about accuracy. Both architectures run comparable algorithms, and the differences that decide a deployment are structural.

What does the bandwidth arithmetic look like?

The cost of cloud face recognition on live video is not the API price, it is the upload. A continuous stream is a continuous commitment, and a site that adds cameras adds bandwidth linearly. Running the model where the frames already are turns that into occasional metadata: a template, an event, a thumbnail.

1. Count the cameras that will be analysed, and the frames per second each one is analysed at - not the frames per second it records at.
2. For a cloud design, that product is the sustained upload the site must carry during business hours, every day, whether or not anything happens.
3. For an on-premise design, it is the load on local hardware, and what leaves the site is the events only.
4. Compare the metered API cost at that volume against the appliance cost amortised over its life, and note where the two cross.

The crossover moves with volume, which is why cloud is often cheaper for a pilot and rarely cheaper for a permanent multi-camera deployment. A comparison run at pilot volume and then applied to the full estate is the most common way this decision goes wrong.

When is a cloud face recognition service the better answer?

Often enough that a page claiming otherwise should not be trusted. Cloud services are the right choice for occasional verification rather than continuous analysis, for teams with no hardware to run, for workloads spread thinly across many locations, and for anything that needs to be live this week.

• Document-to-selfie verification during onboarding: a handful of comparisons per user, no video, no gallery.
• A proof of concept where the goal is to learn whether recognition helps at all.
• An organisation with no place to put a server and no one to look after it.
• A workload that has to scale from nothing to a spike and back, where paying per request is the point.

One caution that applies whichever way the decision goes: identification against a gallery is restricted or unavailable on several major cloud face services, while verification is not. Check that the mode you need is offered before comparing anything else.

What should be evaluated once the architecture is chosen?

• A pilot on the site's own cameras, reporting both error types rather than only successful matches.
• Where templates are stored, and what a breach of that store would expose.
• What happens on a network or power failure, including whether doors fail open or fail secure.
• Which independent evaluation the algorithm has been through, by algorithm identifier and date.
• Whether liveness has been tested against named presentation attack instruments, and by which laboratory.

Independent evaluation results come from NIST, which measures submitted algorithms on sequestered data and publishes what it finds. NIST does not endorse vendors and issues no certification, so a result narrows a shortlist rather than settling it; the pilot measures the installation, which is the number a site actually operates against.