Computer vision on the shop floor: what lighting does to accuracy

A vision model that works in a demo often struggles in week one of a real deployment. The usual reason is not the model — it is the light.

Collect footage across the day

We gather recordings from morning, afternoon, night and the hours when staff restock, so the model sees every condition it will face.

Tune per site, not once

Store layouts, ceiling heights and lamp types differ. A small amount of per-site tuning usually beats a single model trained on everything.

“The camera sees the store as it is at 3 a.m., not as it looked in the demo.”

Measure in the field

We report accuracy on footage from each real camera, not on a test set, and keep doing so after launch.

Plan for the light before you plan for the model. It saves weeks of disappointing results after launch.