Computer Vision on the Retail Shelf: From Photo to Planogram Compliance
How image recognition turns a field rep's photo into out-of-stock alerts, facing checks and trade promotion insights, and what it takes to reach 85% accuracy across hundreds of SKUs.
By David Kukharchuk
Tech Lead at Mirko
The problem with manual audits
Brands spend a large share of field time photographing shelves and filling forms. The data arrives late, is inconsistent between reps and rarely leads to action the same day. Computer vision changes the loop: a photo is analysed in seconds and the result lands in the CRM the rep is already using.
What the model has to recognise
A production shelf recognition system covers
- Individual SKUs, including near-identical variants of the same brand.
- Facings and their order against the planogram.
- Out-of-stocks, gaps and misplaced items.
- Price tags and promotional materials.
In our tobacco recognition project the model tracks more than 200 SKUs with 85% accuracy on decent photos and processes an image in 2 to 4 seconds. The integration with Salesforce matters as much as the accuracy: insights only count when they reach the people who can fix the shelf.
Getting the data right
The hardest part is not the architecture but the dataset. Real store photos have reflections, odd angles and partial occlusion. We collect training images from the field, label them with the client's category team and retrain as packaging changes. Accuracy per SKU is a metric we report, not a single average.
Business impact
Automated audits save up to 90% of the time spent on in-house checks, reduce field headcount pressure and give category managers a daily picture of on-shelf availability instead of a monthly one. That is the difference between a nice model and a merchandising tool.
