How Shelf-Monitoring Computer Vision Is Reducing Out-of-Stocks for a Grocery Chain

The Challenge

Inventory records said a product was in the store, but that did not mean it was on the shelf where shoppers could find it.

The existing approach made it difficult to:

• Spot empty or low facings between scheduled shelf audits.
• Catch out-of-stocks during evening peaks and weekends.
• Tell a true out-of-stock apart from stock sitting in the backroom.
• Check planogram compliance and price-label accuracy across hundreds of aisles.
• Prioritize restocking by sales impact rather than by whoever noticed first.
• Give category and replenishment teams store-level evidence of shelf conditions.

From a technical perspective, grocery shelves are hard for deep learning models. Thousands of SKUs look alike, packaging changes with seasonal promotions, lighting varies from aisle to aisle, and shoppers and carts block the view. Detections had to map to the right SKU, shelf position, and planogram, and the system had to work with the chain's existing inventory and task tools rather than adding another dashboard.

The Solution

InnovationM developed a shelf intelligence layer that turns shelf images into structured availability events and store tasks. Each image is matched against the planogram and current inventory, so the system knows what should be on the shelf, what is there, and what needs action. The solution included:

• Gap & low-stock detection: Using object detection to find empty facings, thin stock, and pushed-back product on each shelf segment.
• SKU recognition: Fine-tuning a product recognition model on the chain's catalog images, with embedding-based lookup so new packaging can be added without full retraining.
• Planogram compliance checks: Comparing detected product positions with the planogram to flag misplaced items and missing shelf labels.
• Price-label reading: Applying OCR to shelf-edge labels to catch missing or mismatched prices against the pricing system, validated through testing services.
• Backroom-aware alerts: Checking on-hand inventory before raising a task, so associates restock items that exist in the store and replenishment is flagged for items that do not.
• Task routing: Pushing restock tasks to associate mobile apps on handhelds, ranked by item velocity and time of day.
• Privacy safeguards: Blurring shoppers and staff in captured images before storage, with short retention windows.

The key focus was not simply detecting empty shelves, but creating a way to send the right associate to the right aisle while the sale could still be saved.

The Impact

The implementation helped the grocery chain move from periodic manual checks toward a continuous shelf signal that store teams act on during the shift. Category teams also gained shared evidence of shelf conditions across stores. The solution supported:

• Fewer out-of-stock hours on high-velocity items.
• Faster time from empty facing to restock.
• Higher planogram compliance.
• Fewer price-label errors at the shelf edge.
• Less time spent on manual shelf audits.
• A foundation for store-level shelf data in replenishment and supplier discussions.

By bringing together shelf image analysis, SKU recognition, inventory checks, and task routing backed by data engineering, InnovationM showed how computer vision development services can make shelf availability something a grocer measures and manages every hour of the day.