98%. 99%. 99.5%.

Put those figures side by side and the highest one looks like the obvious winner.

But the percentage is only one part of the result. A test covering a small number of visitors over a short period does not provide the same evidence as one covering thousands of visitors across changing traffic conditions. Sample size, duration and statistical confidence matter alongside the percentage.

Scope matters too. The same figure can describe a laboratory best case, one reference location, an average across hundreds of sites or a performance floor across almost every deployment. Those are different claims.

Before comparing any people counting accuracy claim, ask six questions:

1. What was measured?

2. How much was measured?

3. Where was it measured?

4. What was used as ground truth?

5. Who designed and ran the test?

6. Where is the evidence?

This guide explains what a complete answer looks like, which red flags to watch for and how Indivd's own accuracy claims stand up to the same test.

A counting error does not stay in the counting report

People counting sits underneath other metrics teams use to evaluate physical locations. If the count is wrong, the error carries into the metrics built on it.

Take a store with 970 recorded visits and 150 transactions. Reported conversion is 15.5%. If 1,000 people actually visited, conversion is 15.0%.

The difference starts with 30 missed visits. It ends up in the conversion rate used to compare that store with others.

The same principle applies to staffing decisions, location comparisons and other analysis that uses traffic as an input.

Illustrative example, not Indivd data.

Two accuracy percentages are rarely measuring the same thing

People counting accuracy describes how an installed system performs under specific conditions. Change the conditions and the result can change.

Mounting height, camera angle, entrance width, lighting and traffic density can all affect counting performance. A system may perform well at one entrance and less well at another, even when the same type of technology is used.

An Indivd benchmark across four home improvement stores illustrates the point. All four used the same type of overhead 3D people counter. One store undercounted every day by 11.7% to 19.2%, while another moved from a 4.4% overcount to a 14.8% undercount.

This means a best result, an average and a performance floor are different claims:

• Best result: What has the system achieved under favourable conditions?

• Average: What does the system typically achieve across the sites included in the sample?

• Performance floor: What minimum level does the claim say the system achieves across a defined share of deployments?

For procurement and operational planning, a performance floor can be more useful than a best result or average. It gives buyers an expectation they can test at their own locations.

How to test a people counting accuracy claim

A percentage only becomes meaningful when you understand how it was produced. Use these six questions to establish the scope behind any people counting accuracy claim before comparing vendors.

1. What was measured?

Start by asking what the accuracy percentage describes.

Depending on the system, an accuracy claim may refer to different measurements, such as:

• entries and exits across a counting line;

• occupancy within a defined area;

• movement between defined zones;

• unique visitor counts;

• visitor journeys through a location.

These are not interchangeable measurements. A system that accurately counts crossings at an entrance has not necessarily demonstrated the same accuracy in distinguishing unique visits or reconstructing visitor journeys.

Ask the vendor to name the exact metric, the unit being compared and how the error was calculated.

Is the figure net or absolute?

This distinction can materially change the percentage.

A net calculation allows errors in opposite directions to offset one another. An overcount in one period can cancel an undercount in another, producing a smaller net error even though both periods were miscounted.

Absolute accuracy treats errors separately, so errors in opposite directions do not cancel.

Two percentages should not be compared until you know how each was calculated.

2. How much was measured?

A percentage needs a sample behind it.

Ask how many visitors, hours, entrances and cameras were included in the test, and whether the sample covered different traffic conditions. A result based on a small number of crossings tells you less about performance across normal operating conditions than a larger sample covering changing traffic volumes.

Where a confidence interval is reported, ask for that too. It tells you how much statistical uncertainty sits behind the headline percentage.

A useful accuracy claim should therefore give you both the percentage and the size of the test behind it.

Indivd's QA report for a fashion store reported a minimum absolute accuracy of 97.78% across 3,101 visitors, 54 hours, 10 counting lines and five cameras. Those details put the accuracy result in context by showing the test behind it.

3. Where was it measured?

Counting conditions vary from one location to another.

A wide, well-lit entrance with an ideally positioned camera is different from a narrow doorway with strong backlight, dense group traffic or limited mounting options. A result from the first site does not predict performance at the second without further evidence.

Find out whether the published figure describes:

• a laboratory test;

• one reference location;

• a selected group of locations;

• an average across deployments;

• a defined performance floor across production sites.

Even within the same retail category, conditions can differ. In Indivd's published accuracy report for a footwear brand, three stores were assessed as separate deployments, with the lowest result at 98.03% absolute counting accuracy. Reporting the stores separately makes the scope visible rather than treating one location as representative of the others.

4. What was used as ground truth?

An accuracy percentage compares the system’s output with a record of what actually happened. That reference is the ground truth.

One ground-truth method is manual review of recorded footage, with crossings reviewed person by person and compared with the system count. Whatever method is used, the vendor should state:

• how the reference count was created;

• who completed it;

• how much footage was reviewed;

• how many visitors were included;

• whether both quiet and peak periods were tested.

"Internal testing" tells you who tested it, but you still need to know how the ground truth was produced.

Verification has its own privacy question

When ground truth is created by manually reviewing real visitor footage, the verification process has its own data-handling requirements.

Ask what state the footage was in when a reviewer saw it, who could access it, how long it existed and what happened to it after the review.

In Indivd's accuracy QA process, QA analysts review full-body masked video fragments rather than identifiable footage. The fragments are automatically deleted after the manual count, while statistical aggregates are produced for the accuracy report.

Learn how Indivd handles data during accuracy verification

5. Who designed and ran the test?

The phrase “independently verified” needs clarification.

An external party can play different roles in a test. It may design the methodology, execute a methodology supplied by the vendor, oversee the evaluation, or perform a combination of those roles.

These roles provide different information about the independence of a result.

Ask who:

• designed the method;

• selected the sample;

• ran the test;

• analysed the results;

• had a commercial interest in the outcome.

A useful claim describes what the external party actually did rather than relying on the word "independent."

6. Where is the evidence?

A published figure should be supported by evidence that can be reviewed later.

Ask what evidence sits behind the published figure and whether the vendor retains:

• the system count;

• the reference count;

• the difference between them;

• the location and camera details;

• the dates, duration and traffic volume;

• the final accuracy calculation.

You should also know how long these records are kept, what exactly is retained and what a customer or third party can review later.

This distinction matters when verification footage itself is deleted after review. Retaining the result is not the same as retaining the original footage for later recounting.

How location affects people counting accuracy

The conditions behind people counting vary across retail stores, shopping centres and airports.

Entrance geometry, camera placement, lighting and traffic patterns change from one physical environment to another. An accuracy result from one setting therefore needs context before it can be applied to another.

The point isn't that one environment is inherently harder than another. An accuracy result only tells you something useful about your location once its test conditions are understood.

Why accuracy should be checked again

Accuracy describes an installed camera under current conditions. It does not describe the product forever.

A camera may be moved during store maintenance, refocused after an installation change or affected by a new fixture near the entrance. Lighting and traffic patterns can also change.

Ask the vendor whether accuracy is checked periodically after installation, what changes trigger a new verification and whether customers can request one when the data looks unexpected.

Applying the test to Indivd's own claims

The same standard should apply to every vendor. Including Indivd.

Indivd reports people counting accuracy as absolute accuracy. Errors in opposite directions are counted separately rather than cancelling one another.

Each QA report states the number of participants and confidence interval alongside the accuracy result, so the percentage can be read together with the size and statistical confidence of the test.

Within that methodology, we make two accuracy statements with different scopes.

The general deployment claim

Indivd's general claim is a performance floor: absolute counting accuracy of 97.5% or better in 99 out of 100 deployments.

The same quality-assurance methodology is applied across deployments. Results vary between sites because camera placement and physical conditions vary.

The customer-commissioned review

The second is a scoped result: approximately 99% absolute accuracy from a customer-commissioned review executed by KPMG.

The review was independent in execution, not methodology. KPMG applied Indivd's quality-assurance method rather than designing a separate testing method. Its scope is one customer-commissioned review.

Re-verification and evidence

Indivd's accuracy QA process includes verification after installation, periodic checks and re-verification following significant changes to camera placement or configuration.

Learn how Indivd verifies data accuracy

The results behind these claims are retained indefinitely and can be checked by a third party.

See the evidence for yourself

A headline percentage tells you little about where and how it was achieved.

Indivd publishes accuracy results from real-world deployments across fashion, food retail, home improvement and other industries, with the test scope and conditions behind each result.

Explore Indivd’s published accuracy reports

A buyer’s accuracy checklist

Before a pilot or contract, ask the vendor to provide the following in writing:

✔ the metric covered by each accuracy claim;
✔ whether the figure is net or absolute;
✔ whether the figure is a best result, average or performance floor;
✔ the number of visitors, hours, entrances and cameras tested;
✔ the statistical confidence reported;
✔ the locations and conditions included;
✔ the ground-truth method;
✔ how verification material was handled;
✔ who designed and executed the test;
✔ what role any independent party played in designing, executing or overseeing the test;
✔ the events that trigger re-verification;
✔ what evidence is retained and for how long;
✔ what a customer or third party can review.

Do not compare percentages until these details are available. Two vendors can publish the same figure while offering very different levels of evidence.

Ask for the evidence, then compare the numbers

People counting accuracy affects every metric built on the count, from store conversion and staffing to shopping-centre traffic analysis, passenger flow and location performance comparisons.

It deserves more scrutiny than a percentage in a comparison table.

The most useful claim is the one you can check.

Look for the vendor that can tell you exactly what was measured, how much was measured, where, against what ground truth, by whom, and then show you the evidence.

That is a standard buyers can apply to every vendor.

Including Indivd.

Common questions