INDUSTRY FIELD GUIDE
Grocery shelf data for AI inventory and product recognition
A grocery aisle changes even when the products stay the same.
NAICS 445110 · Supermarkets and Other Grocery Retailers (except Convenience Retailers)
A grocery aisle changes even when the products stay the same. Packaging refreshes, shelf labels drift and one flavor can block the view of another. That makes a sequence of shelf conditions more useful for evaluating recognition than a folder of clean product shots. Figure has publicly named retail shelf stocking in its contributor program. GS1’s GTIN standard provides a way to reason about when a product or packaging level requires a distinct identifier. Neither establishes demand for a particular store’s footage. Figure Index · GS1 GTIN Management Standard
Record the discrepancy, then the correction
Build one permissioned display or use a closed training aisle. Create a planogram excerpt and use dummy labels or approved product identifiers. Capture a short sweep with one out-of-stock position, one facing of a lookalike item in the wrong position and one shelf-label mismatch. Keep the verified shelf state as a separate record. Then record the restock or correction and label what actually changed.
Provide frame-level examples for recognition or a timestamped event log for a stocking sequence. For a blind evaluation, hold out product families or package designs. Images taken seconds apart can make a test appear better without measuring performance on new products.
Protect shoppers and business records
Ask the store to approve the capture area, labels and time window. Keep customers and bystanders outside the collection. Review retailer and brand rights, store layout, displayed prices, confidential promotion plans and any inventory system exports separately. Avoid using CCTV or point-of-sale records for the first pilot.
A first brief can be one shelf. State the proposed product group, permission, expected output and whether you are offering approved images or a new capture session. Describe your dataset
Original Fiund exercise
Separate a shelf's expected item from its observed item
An original closed-display exercise uses dummy products, labels and a small practice planogram. No shoppers, actual pricing or store system exports appear.
| Record | Source | What it records |
|---|---|---|
| PLAN-G1 | Practice planogram | Slot S1 is assigned product Alder. Slot S2 is assigned Birch. |
| VIEW-G2 | Verified practice shelf sweep | S1's readable package is Birch. S2 is empty. The shelf-edge label at S1 still says Alder. |
| RESTOCK-G3 | Recorded correction | Birch is removed from S1 and Alder is visibly placed there. No S2 replacement is shown. |
| LEDGER-G4 | Separate inventory entry | An entry claims both slots were replenished, without another supporting view. |
Question for the model
What discrepancies and corrections are supported by the sweep, and does it confirm G4's full claim?
Show the reference answer and checks
The initial sweep shows a wrong item in S1 and an empty S2. G3 documents S1's visible correction to Alder. It does not show S2 replenished, so the video does not confirm the ledger's full claim. Keep expected assignment, observed package, edge label and entered inventory as separate sources. A matching shelf label alone would not identify the item in front of it.
- Using a shelf-edge label as proof of the package actually present.
- Treating an inventory entry as an observed correction for an unshown slot.
Fields to preserve
- Slot and expected item
- Retain the planogram's reference state.
- Observed package identity
- Record what is readable in the view.
- Correction interval
- Identify exactly which state changes are visible.
- Inventory-event source
- Keep entered values separate from visual confirmation.
Test another case
Use a new package design with the same readable dummy identifier and a relocated shelf label. Test source reading rather than appearance memorization.
Use an approved closed area and original props. Exclude shoppers, CCTV, receipts, confidential layouts, promotion plans and real point-of-sale data.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.