# Shelf Monitoring with Computer Vision: Detecting Out-of-Stock Items

> How shelf monitoring with computer vision spots empty facings and planogram gaps, which camera setup suits your stores, and how to run a fair pilot.

- URL: https://weaveward.com/blog/shelf-monitoring-out-of-stock-detection
- Published: 2026-04-28
- Updated: 2026-04-28
- Category: Computer Vision
- Tags: shelf monitoring, out-of-stock detection, retail computer vision, planogram compliance
- Author: Weaveward Team

An empty shelf does not show up in your sales report. When a shopper cannot find their usual cooking oil or instant noodles, they buy a competitor's product or walk out, and the system simply records nothing. Meanwhile your inventory screen may say the item is in stock, because the cartons are sitting in the backroom.

Shelf monitoring with computer vision closes that gap. Cameras or staff phones capture shelf images, models detect empty spaces and recognize products, and the system alerts staff when something is missing, misplaced or wrongly priced. It works, but results depend heavily on how images are captured, how similar your products look, and whether someone acts on each alert.

This guide explains what these systems detect, the four ways to capture shelf images, how the vision pipeline works, and how to run a pilot that ends in a clear go or no-go decision.

## What a shelf monitoring system detects

- **Out-of-stock gaps.** Empty facings where a product should be. This is the most valuable signal and the easiest to detect, because an empty space looks very different from a product
- **Low stock.** Only one or two units left, or a facing that looks full but has nothing behind the front row. Shelf depth is hard to judge from one camera angle, so treat this as an estimate
- **Planogram compliance.** Whether the right products sit in the agreed positions with the agreed number of facings. Brand principals often pay for promotional displays and want proof that the display is in place, so compliance reports are useful to both retailer and supplier
- **Price label mismatches.** Reading shelf labels with OCR and comparing them with the POS price, so a promo price that ended last week does not stay on the shelf
- **Misplaced items.** Products put back on the wrong shelf by shoppers, or a competitor's product sitting in a paid display

### The phantom inventory problem

Many out-of-stock situations are not supply problems. The stock exists in the store, but it is in the backroom, in the wrong aisle, or miscounted in the system. This is often called phantom inventory, and it is why shelf monitoring pays off most when combined with stock data.

If the camera sees an empty shelf while the inventory record shows 40 units, the right action is "replenish from the backroom" or "recount this item", not "reorder". Our guide to [inventory management system features](/blog/inventory-management-system-features) covers the stock data side.

## Four ways to capture shelf images

| Method | How it works | Strengths | Limitations | Fits best |
|---|---|---|---|---|
| Fixed shelf cameras | Small cameras on opposite shelves or ceilings take photos at set intervals | Continuous, no staff time | Hardware, power and network per aisle; shoppers block views | High-value categories in large stores |
| Existing CCTV | Reuse security cameras | No new hardware | Angles and resolution rarely show shelf detail | Large gaps in wide aisles only |
| Staff phone audits | An app guides staff to photograph each bay on a schedule | Low cost, works in any store | Depends on discipline; not continuous | Minimarkets, many small stores, merchandiser teams |
| Robots or scanning carts | Devices travel the aisles and capture images | Wide coverage, consistent images | High cost; needs suitable layouts | Hypermarkets and large formats |

Existing CCTV is tempting because it is already paid for, but security cameras are placed to watch people and entrances, not to read product labels. Our article on [retail CCTV analytics](/blog/retail-cctv-analytics) explains what those cameras do well.

There is also a camera-free first step. If an item that normally sells every hour has not sold for half a day, it may be off the shelf. Sales-based detection costs little, works across all stores at once, and uses the same principles as [anomaly detection](/blog/anomaly-detection-fraud). It only works for fast-moving items, but it can prove the value of faster replenishment before you buy any hardware.

## How the computer vision pipeline works

1. **Capture** images from cameras or the staff app, tagged with store, aisle and bay
2. **Check quality.** Reject blurred, dark or glare-heavy images and ask staff to retake them while they are still at the shelf
3. **Detect shelves and products.** An object detection model finds shelf rows, individual products, price labels and empty spaces. Our explainer on [object detection](/blog/object-detection-explained) covers how these models are trained and deployed
4. **Recognize each product** at SKU level by comparing it with reference images
5. **Compare with the expected layout,** taken from a planogram or from how the bay looked when fully stocked
6. **Raise alerts** with a photo crop, product name, location and suggested action

### Why product recognition is the hard part

Detecting an empty gap is relatively easy. Knowing exactly which product is missing is harder. A supermarket carries thousands of SKUs, many with near-identical packaging: the same snack in three flavors, the same shampoo in two bottle sizes, or a special Ramadan pack that appears for a few weeks. Packaging also changes without notice.

A practical design uses recognition by similarity. The model turns each product image into a numerical fingerprint and finds the closest match in a reference library, so adding a new SKU means adding a few reference photos rather than retraining the whole model. For products that remain ambiguous, the system can read text on the pack or use shelf position as a hint.

### Edge or cloud processing

Shelf images are large, and store internet connections vary. Running detection on a small device in the store and sending only results and image crops reduces bandwidth and keeps alerts working during outages. Our guide to [edge AI for computer vision](/blog/edge-ai-computer-vision) covers the hardware trade-offs.

Whichever you choose, shoppers will appear in some frames. Blur faces, keep retention short, and treat identifiable images as personal data under Indonesia's Personal Data Protection Law.

## Turning alerts into replenishment

Detection without action changes nothing. Decide these points before the first camera goes up:

- **Who receives alerts:** the staff member responsible for that aisle, through a simple mobile app, not a busy group chat where alerts get buried
- **How alerts are prioritized:** by sales velocity and margin, so a gap in a best-selling item comes before a slow mover
- **How alerts close:** staff mark them as fixed, optionally with a photo, or flag "no stock in backroom", which escalates to purchasing
- **What managers see:** a dashboard of open gaps, time to fix and repeat problem items by store and category

Measure outcomes, not just detections: on-shelf availability from periodic manual audits, average time from alert to fix, and the share of alerts that turned out to be false.

## Running a pilot that gives a clear answer

A pilot should answer two questions: is detection accurate enough in your stores, and does replenishment actually get faster?

1. **Choose 2 to 5 stores and 1 to 3 categories,** ideally fast-moving categories where gaps hurt sales
2. **Measure a baseline** with manual shelf audits at fixed times for two weeks
3. **Install capture and collect reference images** for every SKU in the pilot categories
4. **Measure detection accuracy** against manual audits: precision (alerts that were real gaps) and recall (real gaps the system found), as explained in our guide to [evaluating AI accuracy](/blog/evaluate-ai-accuracy-before-deployment).
5. **Measure the operational change:** time to fix and on-shelf availability compared with the baseline
6. **Decide** whether to expand, adjust the capture method, or stop

As a rough planning range, assuming access to the pilot stores and clean product master data, a pilot of this shape takes about 8 to 12 weeks from site survey to decision. The main drivers of that range are hardware installation and collecting reference images, rather than model training.

Watch for common pitfalls: piloting only in your best-run store, changing staff routines at the same time without tracking it, and judging the system on a vendor demo set instead of your own shelves. For other retail use cases, see our [retail industry page](/industries/retail).

## How Weaveward can build this for you

We would start with a site survey of a few stores to choose the capture method, then build the pipeline for your pilot categories: image capture through cameras or a staff app, detection and product recognition models trained on your own shelves, and an alert app connected to your inventory data. The first deliverable would be a working pilot with an accuracy report measured against manual audits.

If the pilot meets the targets you set, we would extend it to more categories and stores and connect it to your replenishment and purchasing workflow. Learn more about our [computer vision services](/services/computer-vision), or [talk to our team](/contact) to plan a pilot.