When the Training Data Is People

Training data is the part of a vision project most likely to be assembled informally and least likely to be documented. When it contains people, that is a problem.

What applies

Images of identifiable people are personal data, and collecting them for training is a distinct purpose from recording for security.

Which means its own basis, its own notice, and its own retention.

Where the images are of employees, the workplace analysis applies, including consultation and the weakness of consent.

A dataset assembled from operational footage without addressing this is a problem carried forward indefinitely, because it is embedded in a model that cannot easily be untrained.

The practices that avoid most of it

Do not collect people where the target is an object. Blur or mask at capture for inspection, counting of objects, and infrastructure applications.

Use overhead views where the task allows, which capture far less identifying detail.

Discard frames after labelling, retaining annotations and derived features rather than images, where the training approach permits.

Use synthetic or public data for the general classes, and your own only for site-specific tuning.

Each of these reduces the obligation by reducing what exists, which is more reliable than managing it.

Where people must be in the data

Be explicit about the purpose in the notice, separately from the recording notice.

Define retention for the dataset, with a review date, and apply it.

Record provenance per item: where it came from, when, under what basis.

Restrict access to the dataset as tightly as to the live system.

Do not let it leave to a vendor without an explicit term covering it.

Composition, again

A dataset drawn from one site reflects that site's population.

A model trained on it performs worse elsewhere and worse for anyone unlike that population, which is the uneven error distribution described earlier arriving through the back door.

Record the composition to the extent you can without collecting sensitive attributes, which is a genuine tension: you cannot easily measure representativeness without recording the characteristics you should not hold.

The practical route is to evaluate on an external benchmark that reports subgroup performance, and to watch operational proxies — override rates, complaints, repeated flags on the same people.

The labellers

Frequently overlooked and it matters.

Where labellers are contracted externally, the images are being transferred, with everything that entails.

Where labellers are staff labelling footage of colleagues, that is a further processing purpose and it needs saying.

Labelling guidance for people-related classes needs care: any class that requires a labeller to judge intent, mood or suspicion imports their assumptions into the model at scale.

If a class cannot be labelled consistently without judging a person's state, it should not be a class, which is the test from the boundary section applied at the data stage.

The masking decision at capture

The single design choice that removes most obligations from object-focused applications.

Where the target is an object, mask people at capture — on the device, before storage.

Retain the detection, not the frame.

This applies to inspection, object counting, infrastructure survey and agriculture, which together are most of the successful deployments in this field.

It is a configuration setting in many products and is off by default.

Record that it is enabled, and verify it by examining stored data rather than trusting the setting.

The question to ask at the start

Asked early, it avoids assembling a dataset you later cannot defend.

Could this model be trained without images of identifiable people?

For inspection, object counting, infrastructure and agriculture: almost always yes.

For person detection: public datasets plus masked local tuning gets close.

For anything requiring faces: no, and that is the point at which the obligations become substantial.

The answer is more often yes than teams assume, because the question is usually not asked until after the footage has been collected.