Computer vision in operations: what to measure before you scale

The short answer

Computer vision can turn the cameras you already have into operational events: a safety breach, a quality defect, a compliance failure, with a snapshot and a clip as evidence. Before scaling, measure it on your own cameras, on model quality and on operational value, and be wary of anyone who promises perfect accuracy.

Detection is not the product

Detecting a person or an object is the input. The product is the event: a person in a controlled zone, without the required equipment, for longer than the threshold, sent to an alert, a system or a person to review. Most of the value, and most of the work, is in that layer of context, time and rules.

What to measure in a PoC

• Precision: of the events raised, how many were real.
• Recall: of the real events, how many were caught.
• False positives and negatives: because each one has a different cost in your operation.
• Latency: how fast the event arrives compared with how fast you need to act.
• Review time: how long a person takes to confirm or reject an event.
• Operational value: incidents prevented and inspection hours removed.

What a serious provider will not promise

Not 100% accuracy, not zero false positives, not that any camera will work, and not that facial recognition is needed. Camera position, resolution and lighting matter, and they should be assessed before any promise is made.

Privacy and deployment

Most operational use cases do not need to know who the person is, only what happened. Anonymous tracking reduces privacy risk. Processing can run at the edge, in the cloud or in a hybrid setup, depending on bandwidth, latency and where the data may legally go.

What a PoC looks like in practice

A typical PoC runs in stages: an operational discovery to agree which event matters and what it costs; an assessment of the cameras and conditions; a set of real footage labelled as ground truth; a thin end-to-end slice that raises events; an evaluation against the ground truth; and a short operational trial with people reviewing the events. Only then is there enough evidence for a scale decision.

Questions for the operations and safety team

• Which incident, if caught earlier, would save the most money or risk?
• How is it caught today, and how late?
• Who would receive an alert, and what would they do with it?
• How many false alarms per day can the team absorb before ignoring them?
• Which cameras cover the area, and are they positioned and lit for it?
• Where may the footage be processed and stored?

The answers decide the design more than the model does.

When to scale

Scale when the PoC shows operational value above the cost of running and reviewing the system, on your cameras, in your conditions. Our own engine is ready for PoCs on that basis; we do not present it as a finished product, and you should not accept that claim from anyone without evidence.