Naming a product from a photo is easy. Naming the exact product, one SKU among a dozen near-identical twins, turned sideways, half-hidden behind a shopper’s basket, is where retail computer vision succeeds or fails. At Focal, that’s the problem we’ve spent years on, and here’s how we approach it.
The shelf is a moving target. A product facing forward at 9 a.m. has been turned sideways by noon, and by evening it may be showing only its back. That matters because the side and back of a pack are where brands look most alike. A family-size and a giant-size bag of the same cereal, or two flavors in the same range, can be nearly impossible to tell apart from those angles.
Recognizing “a bag of cereal” is trivial. Knowing the exact SKU, the one barcode among a dozen near-identical twins, is the real challenge, and it’s where the value lives.
We read the whole shelf, not just the product
Our approach starts from a simple belief: a single product image rarely contains enough information to be certain. So instead of judging a product in isolation, our system reads everything the image offers: the product, yes, but also the price tags, the neighboring products, and exactly where the item sits on the shelf. It detects the full scene, including promotions, obstructions, furniture and much more.
The product image is where recognition starts, not where it ends. When the picture alone can’t settle the question, the system looks for corroboration elsewhere on the shelf: the price tag, the neighboring products, where the item sits, and more.
Think of it as a panel of experts rather than a single judge. No one source of information decides the result. A guess becomes a confident identification only when the evidence lines up.
Take a family-size and a jumbo-size bag of the same chip brand, genuinely hard to tell apart from most camera angles. On the image alone, it’s close to a coin flip. But the rest of the shelf isn’t ambiguous: the price tag matches only one of the two, and the surrounding products confirm which section this is. Taken together, the answer is clear, even though the photo alone could never have told you.
That level of confidence is what powers the use cases our customers actually buy: on-shelf availability, out-of-stock alerts, planogram compliance, price compliance, promotion compliance, phantom stock and others.
Where we are genuinely different
I’ll be candid: the difference is not a magic model or algorithm. Plenty of vendors run capable models. Most treat recognition as a single computer vision problem: detect the product, classify it, done. We treat it as a data problem that sits on top of a vision problem.
That starts with the input. Our cameras install quickly and watch the same shelf from the same position, hour after hour. Consistent input is what keeps a model reliable at scale, and it means our system understands each image in the context of the whole store, not as an isolated snapshot.
It continues with the data behind the model. We invest heavily in data quality: identifying missing products, validating product images, collecting multiple viewpoints, removing poor-quality examples, and continuously retraining on the edge cases the system meets in production. In our experience, small improvements to the training data often deliver bigger gains than replacing the underlying model.
The result is a recognition pipeline that gets more accurate over time, instead of drifting as products and stores evolve.
It works on day one, then we tailor it to your store
A large share of most grocery assortments is national brands, and the same cereal box looks the same in every store that sells it. So our foundation model is general enough to start working the moment the cameras go live, trained on more than two billion labeled images from over 300,000 deployed shelf cameras. Dedicated training then closes the gap on what’s unique to you: private label, your price tags, your fixtures.
We believe this is the right order: generalize first, then customize. A retailer operating across several countries shouldn’t need a separate model per country, and every rollout should build on a proven foundation rather than start from scratch.
Our business layer is just as configurable, because retailers define “compliant” differently. Some only require the right shelf and a price tag; others demand the exact facing. We adapt to your definition rather than forcing one rigid standard.
Fresh categories like produce and bakery are a different problem altogether. There’s no packaging to read, so the product has to be judged on appearance alone: variety, ripeness, quality and quantity. That’s why it runs on a dedicated Produce Vision model, purpose-built rather than stretched.
Why this matters for your business
Our customers see this in the numbers that matter to their business. Continuous monitoring has cut manual scanning costs by up to 50 percent and lifted sales by 3 to 5 percent against baseline through better availability.
My advice to any retailer evaluating computer vision is to look past the accuracy headline and ask how a system reaches and keeps that confidence across every store and every hour. Recognizing the shelf the way your customer sees it is what turns a clever demo into a system you can rely on. That is the standard we hold ourselves to, and the standard I think the industry should hold every vendor to.
Frequently asked questions
What is product recognition in retail?
Retail product recognition uses computer vision to identify the exact SKU of every item on a shelf from an image. The useful version does more than name a category. It distinguishes a specific product from visually near-identical variants of the same brand, which is what enables on-shelf availability, out-of-stock alerts, planogram and price compliance and more.
How does AI tell the difference between two nearly identical products?
By looking beyond the product image. A family-size and a jumbo-size bag of the same brand can be hard to tell apart visually, so the system also draws on the rest of the shelf, such as the price tag and the surrounding products. That wider evidence settles what the photograph alone cannot.
What is crowd recognition?
Crowd recognition is Focal’s approach of combining multiple signals from the shelf into a single prediction, rather than relying on one model’s guess. No single source of information decides the result. The system commits to an answer only when the evidence agrees.
Why isn’t a product image enough to identify a SKU?
Because the shelf is a moving target. Products get turned sideways or backwards and get blocked by shoppers and stock. Within a brand line, the sides and backs of different sizes and flavors look nearly the same. A single image often doesn’t contain enough information to be certain, so reliable systems check it against the rest of the scene.
How accurate is shelf product recognition on day one?
It depends largely on how much of the assortment the foundation model already knows. National brands look the same across retailers, so a model pretrained on them can deliver meaningful recognition immediately. The remainder, such as private label, price tags, fixtures and local assortment, is closed through targeted training after deployment.
Can one recognition model work across multiple retailers and countries?
Yes, and it should. National-brand products look the same wherever they’re sold, so a foundation model can learn those patterns once and generalize. Customization is then focused on what is genuinely retailer-specific. Rebuilding a separate model per customer or per country adds cost without adding accuracy.
How is produce recognition different from packaged product recognition?
In most cases, produce and bakery have no packaging to read, so recognition has to judge the product itself from appearance alone: variety, ripeness, quality and quantity. That’s a fundamentally different problem from separating near-identical packaged SKUs, which is why it runs on a dedicated Produce Vision model.
What business results does continuous shelf monitoring deliver?
Focal customers have cut manual scanning costs by up to 50 percent and lifted sales by 3 to 5 percent against baseline through improved on-shelf availability. The gains come from catching gaps and compliance issues throughout the day, rather than whenever a manual audit happens to take place.
