Why It Works in the Demo

Almost every deployment underperforms its demonstration. The reasons are consistent enough to enumerate, which means they can be checked before purchase rather than discovered after.

What a demonstration is

Footage the vendor chose, or a live scene arranged for the purpose.

Good conditions: even lighting, favourable angle, subjects at a helpful distance.

Common classes that the model was trained on extensively.

A threshold tuned to look good, which usually means low false alarms on a short clip.

No fatigue. Nobody watched it for eight hours.

None of this is dishonest. It is what a demonstration is. The mistake is treating it as evidence about your site.

What changes in deployment

Camera placement. Existing cameras were installed for human review at wide angles, not for detection at a specific scale.

Lighting. Dawn, dusk, glare, shadow, headlights, seasonal sun angle. Most sites have several hours a day of hard conditions.

Weather, outdoors: rain on the lens, fog, snow, wind moving vegetation.

Occlusion. People behind objects, behind each other, partially in frame.

Scale. A person at forty metres is a handful of pixels.

Class variation. Your forklifts, uniforms, packaging and signage differ from the training data.

Duration. A detector that produces one false alarm per hour is charming in a demo and unusable across sixty cameras.

The multiplication problem

Rates that seem small multiply by camera count and hours.

One false alarm per camera per day across sixty cameras is sixty alerts a day, which is an alert every eight minutes across a shift.

Compute this before purchase, using the vendor's own figures. It is arithmetic, it takes a minute, and it frequently ends the conversation.

What to ask for instead of a demo

A pilot on your cameras, in your conditions, over a period that includes the difficult hours.

Ground truth collected independently by someone recording what actually happened.

Performance reported by condition: by hour of day, by weather, by camera.

Subgroup performance, where people are involved.

The alert volume per operator per shift that results at the chosen threshold.

The honest expectation

A well-chosen system on well-placed cameras solving a well-defined detection problem works.

The same system asked to detect a rare, visually subtle event across existing wide-angle cameras in variable light will not, and no amount of tuning will fix a camera pointed at the wrong thing.

Which means camera placement is usually the highest-return investment, and it is the one that never appears in the vendor's proposal.

Asking for a reference site

Twenty minutes that is worth more than any demonstration and is routinely skipped.

Ask for a customer with comparable conditions: similar cameras, similar environment, similar scale.

Speak to the operator, not the sponsor.

Ask the false alarm rate per shift, and whether anyone still reads the alerts.

Ask what camera work was needed that was not in the original quote.

Ask what they would do differently.

A vendor unable to offer a comparable reference is telling you something about the deployment base.

The arithmetic in the room

One calculation, done during the demonstration, that changes many decisions.

Take their quoted false alarm figure.

Multiply by your camera count.

Divide by shift hours.

Say the result aloud: an alert every N minutes, per operator.

Ask whether that is workable, and let the room answer.

It takes a minute, it uses their own numbers, and it moves the conversation from capability to operability — which is where it should have started.