Ask most people what product recognition does and they’ll tell you it looks at an item and names it. That is the easy version of the problem, and in my experience it is not the version that decides whether a retailer succeeds with computer vision. At Focal we have spent years on the hard version, and I want to share how we think about it.
The shelf is a moving target. What a camera sees changes hour to hour. A product facing forward at 9 a.m. is turned sideways by noon, showing only its back — and within the same brand line, the back and sides of a family-size bag and a giant-size bag look almost identical. The flavor changes; the packaging barely does. Recognizing “a bag of cereal” is trivial. Knowing the exact SKU, the precise barcode for that one cereal among a dozen near-twins, is the real challenge, and it is 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 beneath it, the neighboring products, and exactly where the item sits on the shelf. It detects the full scene, including promotions, obstructions, and products still sold inside their boxes.
That turns recognition from a one-variable guess — “is this A or B” — into a multivariable decision. Each of those layers casts a kind of vote about what it is seeing, and each vote carries a different weight depending on how much we can trust it in that moment. We combine all of those weighted votes into a single confident answer. We call this crowd recognition, because no one signal decides the result. The consensus does.
At its core, every detected product is converted into a mathematical representation — a compact feature vector that captures its visual characteristics. That representation is compared against millions of known products to generate a shortlist of possible matches.
But we don’t stop there. Each candidate is then evaluated using additional evidence from the scene. Does the price tag match? Is the product located where the planogram expects it? Do neighboring products belong to the same category? Have previous scans seen the same product in that position?
Think of it as a panel of experts rather than a single judge. Every source of information casts a vote, and each vote carries a different weight depending on how reliable it is in that particular image. The system combines those weighted votes into one final prediction.
Take a family-size and a jumbo-size bag of the same chip brand — genuinely ambiguous from most camera angles. The visual vote might come back close to a tie. But the other layers won’t: the price tag isn’t ambiguous, it matches only one of the two, and the neighboring-product vote confirms which shelf section this is. Weighted together, the tie breaks decisively, even though the product photo alone could never have told you which one it was.
On top of that sits what we call a business layer, which adds real-world context. If the system is torn between a bag of rice and a bag of cereal, the surroundings break the tie: you are in the rice aisle, the neighbors are all rice, so it is almost certainly rice. That extra layer of context is what pushes confidence high enough to power the use cases our customers actually buy — on-shelf availability, out-of-stock alerts, planogram compliance, price compliance, and promotion compliance.
Where we are genuinely different
I’ll be candid. The difference is not a magic algorithm. Plenty of vendors run capable models. The difference is what you do with the data. Standard approaches tend to take the image as it is. Our edge is the voting model and the work of tuning the right weight for each layer — knowing when a price tag should count for more and a product image for less — plus training specifically to tell near-identical packs apart, where the small differences are everything. It is also the end-to-end package. Recognition is only as good as the data feeding it, and that starts with cameras that install quickly and watch the same shelf from the same position every hour. Consistent input is what keeps a model reliable at scale.
In our experience, the biggest gains rarely come from changing the recognition model itself. They come from improving the quality of the information that reaches it. Most vendors treat this as one computer vision problem: detect it, classify it, done. We treat it as a data problem sitting on top of a vision problem — the model only ever sees one shelf position at a time, but our system carries context across the whole store: which planogram section a camera covers, what is supposed to be there, and what was there an hour ago. That is what lets us weight one variable more heavily in one context and a different variable more heavily in another, instead of applying one static threshold everywhere.
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 encounters in production. Small improvements to the training data often produce larger gains than replacing the underlying model.
The result is a recognition pipeline that becomes more accurate over time instead of gradually drifting as products and stores evolve.
It works on day one, then we tailor it to your store
Roughly 80 to 90 percent of a typical retailer’s assortment is national brands. The same Kellogg’s box sits in Whole Foods and Albertsons alike. Because of that, our foundation model is general enough to start working the moment the cameras go live, trained across more than two billion labeled images from over 300,000 deployed cameras. Dedicated training then closes the gap on what is unique to you: private label, your specific price tags, your store fixtures. 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.
We believe the foundation model should generalize across retailers rather than being rebuilt for each customer.
Most grocery assortments are made up of national brands, and within a market those products look exactly the same regardless of which retailer sells them. The model should learn those common visual patterns once, then focus its customization on what is truly unique to each retailer: private-label products, price tags, store fixtures and business rules.
That distinction matters more the larger you get. A retailer operating across several countries should not need a separate model per country — the shared assortment is shared, and every additional store that comes online strengthens the same foundation rather than starting a new one. Generalize first, then customize, is what makes a rollout compound instead of restart.
And categories like produce and bakery, where a banana is a banana and freshness matters more than a barcode, run on a dedicated produce vision model built for that reality. There is no packaging to read and no price tag attached to the item, so the model has to judge the product itself — variety, ripeness, quality and quantity — from appearance alone. That is a different problem from packaged-goods recognition, and it deserves a purpose-built model rather than a stretched one.
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 is the use of 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 barcode from visually near-identical variants of the same brand, which is what enables on-shelf availability, out-of-stock alerts, and planogram and price compliance.
Why do computer vision systems that score 99% in a pilot fail in production?
Because a pilot accuracy score describes the test conditions, not the system. Pilots use controlled lighting, a limited SKU range, standardized packaging and predictable layouts. In production, SKU proliferation, lighting variability, inconsistent store layouts and non-standard packaging occur simultaneously and compound, so a model meets a materially harder problem than the one it was measured on.
How does AI tell the difference between two nearly identical products?
By using evidence beyond the product image. A family-size and a jumbo-size bag of the same brand can be visually ambiguous, so the system also weighs the price tag beneath the item, its shelf position relative to the planogram, the neighboring products, and what previous scans saw in that position. Those weighted signals resolve the tie the photograph alone cannot.
What is crowd recognition?
Crowd recognition is Focal’s approach of combining multiple independent signals from a shelf image into a single prediction rather than relying on one model’s guess. The product image, price tag, shelf position, neighboring items, promotions and historical observations each cast a weighted vote, and the system commits only once the combined confidence score clears a threshold.
Why isn’t a product image enough to identify a SKU?
Because the shelf is a moving target. Products get turned sideways or backwards, get occluded by shoppers and stock, and within a brand line the back and sides of different sizes and flavors look nearly the same. A single image frequently does not contain enough information to be certain, so reliable systems corroborate it against the rest of the scene.
How accurate is shelf product recognition on day one?
It depends almost entirely on how much of the assortment the foundation model already knows. Roughly 80 to 90 percent of a typical grocery assortment is national brands that look the same across retailers, so a model pretrained on that base can deliver meaningful recognition immediately. The remainder — private label, proprietary 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. Within a market, national-brand products look identical regardless of which retailer sells them, so the foundation model can learn those patterns once and generalize. Customization is then applied only to 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?
Produce and bakery have no packaging to read and no barcode on the item, so recognition has to judge the product itself — variety, ripeness, quality and quantity — from appearance alone. That is a fundamentally different problem from separating near-identical packaged SKUs, which is why it runs on a dedicated produce vision model rather than a general one.
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 continuously throughout the day rather than at whatever moment a manual audit happens to occur.
