What 2,400 labelled images of supermarket shelves taught me

When I started Retail Shelf Intelligence, I assumed the interesting work was the model. I was wrong. The interesting work — and 90% of the effort — was the data.

The data is the model

Anyone can download a YOLO checkpoint. Few can say exactly what's on a specific shelf, in a specific store, under specific lighting. A shelf detection system is essentially a dataset with a wrapper around it.

The facts from my own project:

• 2,400+ images collected and labelled by hand.
• ~40% of my time went to preprocessing and relabelling, not training.
• Every accuracy gain I saw came from fixing labels, not from changing the architecture.

What made it painful

Glare. Packaged goods are glossy. Reflections read as edges to a detector. I learned to treat glare as a data problem, not a model problem: more angles, better preprocessing, and honest labels for partially visible products.

Class imbalance. Three products accounted for most of the data and almost all of the easy wins. The long tail of niche SKUs is where the system actually failed — and where nobody wants to spend labelling effort.

Boundary cases. A product half-stocked looks "fine" to a detector and "empty" to a manager. If your labels can't distinguish these, your model won't either.

The rule I'd teach anyone

Label the boundaries, not the easy cases.

An hour spent labelling the confusing 5% of images moves accuracy more than five hours on the easy 95%. That's counterintuitive, because the easy 95% is where progress feels fastest.

Why this matters beyond retail

Every ML project I've touched has the same shape: the model is the smallest, best-understood piece. The system — data quality, evaluation, deployment, monitoring — is where real engineering happens. Labelling those shelves was the best possible reminder of that.

I keep the full pipeline notes in the project documentation.