AI · Computer vision
Computer vision for retail and brands: what a pilot really needs
Telling a genuine handbag from a fake by its photos, checking that your products are on the shelf, catching a crooked label on the line: these are practical jobs for a vision model now. The hard part is rarely the model. It's the images, the labels and the edge cases, and here is how to prepare them.

Three jobs that work today
Authenticity checks from photos. A customer, reseller or support agent photographs the product: stitching, logo placement, the serial label, the packaging seal. The model compares the photos with genuine references and returns "looks genuine", "suspicious" or "need another photo". People review the suspicious ones.
Shelf and listing monitoring. In stores, photos from field reps show whether your products are on the shelf, in the agreed number of facings, with the right price tags, and where the gaps are. Online, the same kind of model checks product images on marketplace listings: is that your current pack, an old version, or someone using your photos to sell something else?
Defect detection. On the line or at the warehouse dock: scratches, dents, missing caps, misaligned labels, wrong print color. A fixed camera and controlled light make this the easiest of the three. Shelf photos from phones are the messiest.
Start with the decision
Before collecting a single image, write down what changes when the model answers, who acts on it, and what each kind of error costs. For authenticity the two errors are not equal. Missing a fake lets one bad item through. Calling a real customer's purchase fake insults them. So the model never tells a customer "fake"; it says "needs a closer look", and a person makes the call.
Define the answers precisely, and always include a third one: "can't tell from this photo". A model allowed to say that is far more useful than one forced to guess.
Then agree on the bar before the pilot starts, in one sentence. For example: "On the test set, the model catches at least nine in ten confirmed fakes and sends no more than one genuine item in twenty to review." A pilot without that sentence ends in an argument about whether it worked.
Images: real conditions, full coverage
- Shoot the way production will shoot. If customers will use phones in bad light, train on phone photos in bad light. Studio pack shots teach the model little about a box with glare on it, photographed on a kitchen counter.
- Cover every variant. A 40-SKU range with three packaging generations is 120 things to recognize, not 40. Add colors, sizes and factory differences.
- Collect confirmed fakes. They are rare, so ask returns, support and your brand protection team for every one they have, with photos.
- Write a photo protocol. Which shots (front, label close-up, serial number, one key detail), at what distance, without flash. A guided capture screen in an app beats "upload a photo".
- Record where each photo came from. Store, phone model, date, batch. When the model fails on a group of images, that is how you find out why.
- Hold some back. Keep a test set the model never sees during training, from different stores, phones and dates. That is where you measure the pilot.
As an illustration, a defect or shelf pilot often starts with a few hundred labeled images per class and adds more where the model struggles.
Labels: who decides what's true
A model learns exactly what its labels say, mistakes included.
- Use people who know the product. For authenticity, that means your authentication experts or QC staff, not a generic labeling crowd.
- Write a labeling guide with example images for every class, including borderline ones.
- Double-label a sample. If two experts disagree on, say, one image in seven, the model can't do better than that. Fix the guide before you train more.
- Mark where, not just what. For defects and shelves, draw boxes around the problem. For authenticity, mark the detail that gave the item away; it also lets the model show the reviewer why it flagged the item.
- Only confirmed cases count as fakes. A suspicion is not a label.
Edge cases to plan for
- new packaging launched in the middle of the pilot;
- genuine variation: different factories, batches, wear on second-hand items;
- photos of screens, photos of printed photos, edited images;
- partial views, blur, reflections, dark corners of a warehouse;
- counterfeiters adapting, so next quarter's fakes look different;
- on shelves: promo displays, a competitor's pack in similar colors, a new price-tag format.
Plan a retraining rhythm from the start, monthly for example, and route every "can't tell" and every reviewer override back into the training set.
Measure the pilot on the held-back images: how often the model is right per class, how many real problems it catches, the share of "can't tell", and time per check against the manual process.
What to do this week
- Pick one job and the decision it changes.
- Collect 50 real photos taken the way they will be taken in production, and look at them as a team.
- Write the classes, including "can't tell".
- Find your experts and your confirmed cases.
- List ten edge cases you already know about.
Boson Agents, the AI lab of Boson Group, built Boson Verify (verify.boson.group) to check product authenticity from photos, and trains its models on its own GPU servers. If you're planning an authenticity, shelf or defect pilot, write to team@boson.group.




