Know what's on every shelf, without sending anyone to check

The Problem

Retail teams lose money to empty shelves every day, and most of those losses stay invisible until someone walks the floor or a customer complains. Today’s tools – manual checklists, planogram audits, monthly store visits – give you a snapshot of yesterday. Real-time visibility is what they lack.

Multiply that by hundreds of SKUs and dozens of stores, and a 30-minute audit becomes a 30-person operation. Most of it is repetitive work, counting facings, checking shelf state, photographing displays. Exactly the kind of task computer vision was built for.

What the Solution Does

A continuous, automated check on the state of every shelf in every store. Cameras capture what’s on the shelf, the system understands what should be there, and the difference becomes an alert your team can act on.

1. Capture, fixed cameras (or existing CCTV) photograph shelves on a schedule.
2. Recognize, computer vision matches each visible product to your catalogue, counts facings.
3. Compare, the system checks recognized state against your planogram and stock levels.
4. Alert, discrepancies (empty shelves, wrong placement) become real-time notifications.
5. Report, historical data feeds dashboards for trend analysis and ops decisions.

The system goes beyond generic object recognition. It tells apart specific beverage brands and individual yogurt flavors. It separates packaging variants of one product, such as 120g and 180g coffee. It also detects new SKUs reliably, even when their packaging looks nearly identical to existing items.

Where It Fits

This makes sense if you…

• Run 20+ retail locations or a large distribution network
• Have a stable SKU catalogue with reference images (or can build one)
• Already have cameras in stores, or can install them
• Currently spend significant ops time on shelf audits / planogram checks
• Need data to negotiate with brands or measure merchandising compliance

This is probably not the right time if you…

• Run a single store or pop-up, manual checks are cheaper
• Have a catalogue that changes every week with no reference images
• Are still finalizing your planogram strategy, the data needs a target to compare against
• Don’t have a clear owner on the ops side who’ll act on alerts

Business Value

The hard ROI shows up in three places.

Operations cost. Once the system covers your store network, staff stop walking shelves with clipboards. Typical reduction in manual inspection time is 60-80%, depending on store size and audit cadence. The people you free up stay on payroll and move to restocking, customer interaction, or higher-value tasks.

Lost sales recovery. Empty shelves stop being invisible. Real-time out-of-stock detection lifts on-shelf availability by around 5-15 percentage points in the first quarter for typical deployments. The exact lift depends heavily on baseline supply chain quality. The system surfaces problems. The team still has to fix them.

Objective insights. Where layout errors happen most often, which products are frequently misplaced, what times violations typically occur. This data lets management make targeted, network-wide decisions grounded in measurement, in place of subjective impressions from store managers.

How It Works

The detection stack uses YOLO-family models for object recognition combined with our own fine-tuning pipeline. Out of the box, the model handles around 95% of routine SKUs after training on your catalogue. As new products appear, we use Datapipe (our open-source ETL tool) to retrain incrementally on new examples, with no rebuild of the whole pipeline.

Camera feeds are processed either on-device (for stores with edge hardware) or in the cloud. Alerts ship to whatever system your ops team already uses, Slack, Teams, your store management app, or a custom dashboard.

Integration with the rest of your stack runs through one API: the system receives the planogram and product catalogue from your ERP/PIM, and returns recognition events and reports. We’ve integrated with SAP, 1C, and custom in-house systems.

Stack

YOLOv8 family, ResNet, Datapipe (ETL), Metabase (analytics), Python/FastAPI, Docker, Kubernetes. Edge: Jetson family. Cloud: AWS / GCP / on-prem.

What You Need to Make This Work

Data. A product master with SKU IDs, names, and 3-5 reference photos per product from typical store angles. If you don’t have reference photos, we run a 1-2 week data collection sprint as part of the pilot.

Integrations. Read access to your planogram source (ERP, PIM, or spreadsheet). An endpoint for delivering alerts (Slack, email, webhook, or push to your store app).

Hardware. Existing CCTV often works. For new installations, expect 1-4 cameras per aisle depending on shelf height and lighting. We help spec the setup during the test stage.

Team. A clear ops owner who’ll act on alerts (this is the most important and most underestimated requirement). Plus a part-time tech contact on your side during integration (~5-10 hours/week for the first month).

Implementation Roadmap

1. Test (2-4 weeks)

Pick one store, or even a single shelf. We install cameras (or repurpose existing ones), train the first version of the model on your top SKUs, and measure baseline metrics: recognition accuracy on bright, small, and similar items, false alert rate, end-to-end latency. Output: a written report on whether the technology works for your conditions, your lighting, your packaging, your store layout, and what would need to change to scale.

2. Pilot (1-2 months)

If the test passes, we roll out to 3-10 stores chosen for diversity (different formats, regions, assortment). During the pilot we tune the model on edge cases and set up the alert workflow with your ops team. We measure the operational metrics that matter beyond detection accuracy: response time, false-alert fatigue, and measurable improvement in shelf state.

3. Scale (3-6 months)

Network-wide rollout with ongoing monitoring, regular model retraining as your assortment changes, and feedback loops to keep accuracy steady. By the end of this phase, your team owns the day-to-day; we stay on for SLA-based support and periodic retraining.

Keep in Mind

The system handles around 95% of routine recognition automatically. The remaining 5% will always need human review. We design for that on purpose.

Cases that typically need manual handling:

• New products without reference images, until we collect a few photos, the system flags them as unknown and holds off on a guess.
• Heavily occluded shelves, when packaging is blocked by other items or staff, confidence drops and the system asks for review, so it does not fire a noisy alert.
• Drastic lighting changes, opening a new window, changing store fixtures, or daylight extremes can require a short retraining cycle.

Accuracy also depends on factors the model can’t fix: camera quality, shelf access, and how consistently your team uploads planogram updates. In typical retail environments, 10-15% deviation is considered normal; large products with bright packaging under good lighting reach 90-95% accuracy, while small similar products under average conditions land at 80-85%. We make these dependencies explicit during the test phase.

The system is a tool. Its job is to save time and make people’s work easier. It does not supervise them. If a planogram is constantly being violated, the problem may not be the staff. It may be the plan itself.

FAQ

How much training data do you need to start?

For a pilot covering your top 100-200 SKUs: 3-5 reference photos per SKU, captured from typical store angles. We can collect this during a 1-2 week data sprint if you don’t already have it.

What happens when we add new products?

The system flags unknown items and holds off on a guess. Our Datapipe pipeline handles incremental retraining, so adding new SKUs takes hours. Older approaches took weeks.

Can this work with our existing CCTV?

Often yes, we evaluate camera resolution, angle, and lighting during the test phase. If existing cameras aren’t sufficient, we recommend a hybrid approach: keep CCTV for security, add dedicated cameras only on critical shelves.

What’s a normal error rate?

10-15% deviation is typical in real retail conditions. The number depends heavily on packaging (bright vs muted), lighting consistency, and the assortment overlap (how many products in one category look near-identical). Accuracy below 80% usually points to equipment or configuration issues. Model limits are rarely the cause.

Ready to Discuss?

If this sounds relevant to your operation, the fastest way to find out whether it’ll work for you is a discovery call. We’ll walk through your current state, number of stores, camera situation, and planogram process, and tell you whether a pilot makes sense, what it would cost, and what to expect.

Discuss your project