A visible shelf gap is evidence, not inventory truth. Learn how to separate camera observations from POS, backroom, and replenishment logic so an out-of-stock workflow can be tested honestly.
A camera pointed at a shelf can see the shelf. It cannot see cases in the stockroom, a product in a shopper’s cart, an item placed elsewhere in the store, a delivery in transit, or a POS record that has not synchronized. Calling every visible gap an out-of-stock event therefore overstates what the vision system knows.
The better design separates an observation from a conclusion. Vision reports a shelf state. Retail data and business rules decide whether that state represents an actionable out-of-stock condition.
| Signal | Meaning | Typical Input |
|---|---|---|
| Gap detected | An expected shelf region appears empty or underfilled | Camera plus shelf region map |
| Low visible count | Fewer visible items or facings than a threshold | Object detection or slot occupancy |
| On-shelf availability issue | The expected product is not sufficiently available to shoppers at that location | Vision plus product and location mapping |
| Out of stock | The retailer determines that sellable inventory is unavailable | Vision, POS, inventory, backroom, and workflow data |
Not every project needs all four. A store may get value from a refill task triggered by a persistent gap even if it never claims to know enterprise inventory.
A durable camera event contains evidence that another system can interpret. At minimum, include device ID, location or region ID, timestamp, detected state, confidence or quality indicator, and model version. Add an evidence image only when privacy policy, bandwidth, and retention rules allow it.
For a fixed-slot shelf, the state could be occupied slots divided by expected slots. For a less structured shelf, it may be a visible count or fill estimate. In both cases, require the condition to persist across more than one observation before raising an operational event.
The application layer maps the visual event to a product, store, aisle, fixture, and task. It can check recent sales, expected planogram position, backroom stock, open replenishment work, and alert history. That context decides whether to create a refill task, request a recount, suppress the alert, or escalate the issue.
This division also makes ownership clear. The camera or edge box can produce structured detections. The retailer or solution provider owns master data, business rules, task routing, dashboards, and inventory reconciliation.
To test this boundary with your own shelf view, send CamThink the proposed observation and event fields.
Start with the smallest output that supports a real action. Empty-slot detection is often easier to validate than open-ended SKU recognition. A fixture with fixed product positions is easier than a mixed shelf with changing layouts. A scheduled image may be enough for slow-changing stock, while continuous video may be needed when the workflow depends on short-lived changes.
The success metric is not the number of detections. It is the number of useful actions created without overwhelming store staff. A production rule should define persistence, cooldown, duplicate suppression, evidence, task owner, and task closure. It should also record why staff dismissed an event.
That feedback exposes whether the problem came from the model, camera view, threshold, store data, or workflow. It is also the evidence needed to decide whether a pilot should expand beyond the first fixture.
