Visual Recognition

How product recognition actually works on retail cameras

Product recognition means identifying which item is in frame, not merely that something is there. A model compares the pixels in a camera frame against a reference catalogue of your own SKUs and returns a best match with a confidence score. That last part matters: the output is a ranked guess with a number attached, not a barcode read.

What it needs from your stores

• A reference catalogue of your assortment. Recognition is relative to a known set. A model trained on someone else's products cannot identify yours, which is why this is configured per retailer rather than shipped ready-made.
• Camera angle and coverage that actually see the facing. Ceiling domes aimed at aisles for security were positioned to record people, not to read shelf edges. Existing cameras are often reusable, but not always in place.
• Re-training when packaging changes. A seasonal redesign or a new size makes an existing SKU look like an unknown one until the catalogue is updated.
• A way to act on the output. A detection nobody reviews is not a control. The routing matters more than the model.

What a detection does not prove

This is where camera-based recognition is most often oversold. A low-confidence or missing match is a signal to look, not a finding:

• Occlusion looks like absence. A shopper, a trolley or a pallet in front of the shelf produces the same empty facing as a genuine out-of-stock.
• Similar packaging confuses models, not just people. Two variants of one brand that differ by a small flash on the label are the standard hard case.
• A gap is not a lost sale. It may be a delayed replenishment, a planogram change or stock sitting in the back room.
• An unscanned item at a till is not theft. It is an exception worth reviewing. Treating the model's output as a conclusion is how these systems generate false accusations.

How it differs from barcode scanning

A barcode is a deterministic read of a label someone deliberately presented. Product recognition is a probabilistic read of whatever happens to be visible, whether or not anyone intended it to be scanned. That is precisely why it is useful for shelf state and for scan-avoidance at self-checkout — those are the cases where nothing was presented to a scanner — and precisely why its output belongs in a review queue rather than in an automated decision.

We publish no headline accuracy figure, because a single number would not be meaningful: detection quality depends on your cameras, shelf geometry, lighting and how visually distinct your packaging is. It is measured per deployment. Loss prevention covers the till-side use of the same recognition stack, and inventory management the shelf-side one.