Shelf Availability Computer Vision

Shelf availability computer vision turns shelf photos into a short list of fixable gaps. The model looks at each bay or fixture image and flags empty facings, products that do not match the planogram, items that have drifted into the wrong location, and price or shelf labels that are missing or unreadable. The goal is not a perfect digital twin of the store. The goal is a practical queue for the store team lead: where to walk, what to check, and what to correct before shoppers hit a blank shelf.

This outcome sits in the stock stage of retail operations and serves a quality purpose. Availability on the shelf is what the customer sees. Inventory systems can say a SKU is on hand while the facing is empty, blocked, or mislabeled. Vision closes that gap between system stock and sellable presentation by reading the shelf as the shopper sees it.

Related work includes Inventory Record Reconciliation when on-hand counts and shelf reality disagree, Planogram Compliance Checker when the issue is layout rather than empty space, and Real-Time Stockout Risk Predictor when you want earlier warning before a facing goes dark.

Useful detection depends on readable shelf imagery and a clear reference for what should be there. Typical inputs include:

• Shelf images from handheld devices, fixed cameras, or robot or associate capture routes
• Fixture, aisle, and bay identifiers so findings map to a walkable location
• Planogram or expected assortment for the fixture (SKU list, facing counts, and positions where available)
• Current store assortment rules (authorized SKUs, temporary voids, promotional sets)
• Optional price or label templates when label presence is in scope

Capture quality matters more than capture volume. Blurry frames, heavy glare, extreme angles, and partial bays produce weak or empty results. Teams that schedule short, consistent capture passes (same height, same distance, same lighting discipline) get fewer false alarms and fewer silent misses.

When shelf images are missing, corrupt, too dark, or otherwise unreadable, the system should return empty output for that fixture. No inferred out-of-stock list. No guessed planogram gaps. An empty result is the honest signal that the model cannot evaluate the shelf, and the store team lead should treat it as a capture or coverage problem, not as “all clear.”

The model does not restock shelves. It flags candidates. A store team lead or floor associate still confirms the finding and completes the fill, face-up, or label fix.

A practical flow looks like this:

1. Capture – Associates or devices photograph assigned bays on a defined route and cadence.
2. Detect – The vision model scores each image for empty facings, planogram mismatches, misplaced items, and missing or unreadable labels.
3. Prioritize – Findings are ranked by aisle, fixture, severity (for example, full void on a high-velocity SKU before a single facing short), and freshness of the image.
4. Review – The team lead opens the queue, optionally checks the annotated image, and accepts or dismisses each flag.
5. Correct – Floor staff pull from backstock, face up existing product, relocate misplaced items, or replace labels. Human confirmation closes the loop.
6. Escalate when needed – If the shelf is empty and backstock is also empty, the issue leaves the shelf-availability queue and joins inventory or replenishment workflows (often via Inventory Record Reconciliation or risk signals from Real-Time Stockout Risk Predictor).

Human-in-the-loop is non-negotiable. Lighting, packaging changes, seasonal wraps, and partial occlusions can look like voids or wrong SKUs. The model’s job is to shorten the search. The store team’s job is to decide and act.

A healthy shelf-availability program is measured by speed and correctness of correction, not by raw detection count.

Focus on:

• Time from capture to first correction on accepted flags
• Accept vs. dismiss rate by aisle and by finding type (void, planogram gap, misplace, label)
• Repeat flags on the same facing after a fill (often a backstock or process issue, not a model issue)
• Empty-output rate by fixture (high empty rates usually mean capture coverage or image quality problems)
• Shopper-facing gaps that still appear between capture cycles (cadence may be too slow for that category)

Separate model noise from store reality. A spike in voids after a delivery delay is operational truth. A spike in dismissals after a packaging refresh is often a reference or model update need. Keep planogram and assortment references current so Planogram Compliance Checker logic and shelf-availability flags stay aligned.

Shelf vision fails in predictable ways. Plan for them explicitly:

• Missing or unreadable images → empty output for that bay; schedule a recapture before drawing conclusions.
• Occlusion (carts, endcaps, promotional dumps) → may hide facings; treat partial views as incomplete, not compliant.
• New packaging or private-label changes → temporary false “misplaced” or “unknown” flags until references update.
• Deep voids vs. thin facings → a single remaining unit can look available to a shopper yet still be a near-stockout; pair with inventory and risk signals when depth matters.
• Backroom vs. shelf → vision only sees the shelf. On-hand in the system does not prove sellable availability until product is worked to the floor.

Do not auto-zero inventory from a single shelf image. Do not auto-order from an unreviewed void flag. Use detections as directed work for people who can see the bay, the backstock, and the shopper impact in context.

Start with a narrow scope: a few high-traffic aisles, a fixed capture route, and a single daily review window owned by the store team lead. Require human accept before any fill task is marked done. Treat empty model output as a data-quality alert. Expand only when accept rates are stable and correction times are short enough that findings are still true when someone arrives at the bay.

Over time, connect accepted voids that cannot be filled from backstock into inventory reconciliation, and use planogram compliance where layout drift is the real problem rather than empty space. Shelf availability computer vision earns its keep when it turns shelf photos into fewer missed gaps and faster, human-confirmed corrections, not when it pretends the store can run without walking the floor.