Choosing cameras for shelf recognition

2024-02-12 · guide

Marha Sebieva – AuthorOlga Tatarinova – Final edit

Most shelf-recognition projects that disappoint were lost at the camera. The model was rarely the problem. A detector cannot recover detail the lens never captured. Before anyone trains anything, five physical choices set the ceiling on accuracy: resolution, field of view, mounting, capture frequency and lighting.

This is the checklist we walk through on every retail shelf-monitoring and price-tag install.

Resolution

It is tempting to assume any camera will do. In practice, below Full HD (1920×1080) the system starts confusing small details, can’t read price tags, and struggles to separate similar products. The smaller the packaging and the longer the shelf, the more resolution you need. A single camera covering several meters of shelving has to spread its pixels across all of it.

Resolution is the wrong place to save money. Cut it and the whole system loses accuracy, from the long shelf to the small print.

Top-down comparison: the same Full HD camera over a 2-3 m section samples each facing densely and the price tag reads, while stretched across several meters the same pixels spread thin and the price tag blurs. — The same Full HD budget. Over a short section each facing gets enough pixels, over a long one the price tag blurs.

Field of view

Field of view is a balance, and both extremes fail.

Too wide, and you get a fisheye effect. Objects near the edges stretch and distort, so products at the ends of the shelf read worse than those in the center. Too narrow, and part of the shelf falls outside the frame. The goal is a lens that shows the whole required area clearly, with no distortion at the edges.

Three field-of-view states against one shelf: too wide distorts the edges, too narrow leaves part of the shelf out of frame, and a matched lens shows the whole shelf sharp to both edges. — Field of view is a balance. Too wide distorts the ends, too narrow drops them, a matched lens shows the whole shelf.

Mounting height and placement

The camera should see the shelf at roughly the angle an average shopper does.

Mount cameras too high and you get a top-down view where products occlude each other. A large part of the assortment becomes invisible to the system. Mount them too low and the upper shelves drift out of focus. Island displays and refrigerators usually need custom mounting, and those setups are worth testing before you commit to a rollout.

Three camera heights against a shelf: at 4.5 m the lower shelves are occluded, at 1.8 m the field of view covers the full shelf, and mounted low the upper shelves fall out of focus. — Mounting height sets the ceiling on what the model can ever see. Eye level wins.

Capture frequency

How often you capture depends on what you are trying to catch. If you don’t need minute-by-minute change, one image an hour is often enough. If you need to catch out-of-stocks the moment they happen, or react quickly to layout changes, you move toward near real-time.

The trade-off is cost. The higher the frequency, the heavier the load on storage and processing. Match the cadence to the decision the data feeds. “As often as possible” is the wrong target.

Lighting: often it matters more than the camera

This is the one teams underestimate most. In bad lighting even a top-end camera is ineffective. An expensive camera in poor light can do worse than a cheap one in good light. And lighting varies across a store, so the same product can look completely different depending on where it sits.

Poor lighting hurts recognition in three distinct ways.

It changes how products look. Colors distort, contrast collapses or blows out, small packaging text blurs. In dim light red can read as brown and blue as near-black, and the system starts confusing similar products. We have seen the same can of cola recognized as two different items, once in a fridge, once on an open shelf, purely because of light.

It creates shadows, overexposure and glare. Upper shelves cast shadows on lower ones. Items deep on a shelf fall into the dark, and the system sees only fragments of a package. Direct light blows highlights into white patches with no detail, common near windows on sunny days. Glass surfaces are the worst offenders. A refrigerator door reflects like a mirror, so the camera captures the reflection and loses the product. The glass can be clean and the product perfectly placed, and the system still sees nothing.

Side elevation: overhead light reaches the front lips, each shelf shades the bay below it, and products set deep on a shelf fall into shadow, so the camera sees only the lit front edges. — Overhead light reaches the front lips. Each shelf shades the bay below, and deep items fall dark.

It pushes the system into guessing. When the image is too poor to read, a weak system starts to assume: “this shelf usually holds product X, so it’s probably X”. That is unreliable. Products move, displays change, layouts get rebuilt. The whole point of recognition is to measure reality. Guessing defeats it.

Four real installations

The fastest way to learn the rules is to see them broken and respected.

Failed: ceiling camera in a hypermarket

A 4K camera at 4.5 m, 120° field of view, one camera covering three double-sided units, about 15 m of shelving.

The result: upper shelves block the lower ones, and products show up as colored stripes with no detail. Only part of the assortment is recognized, and price tags are unreadable. The cause was trying to save on camera count. Covering too large an area from a single viewpoint destroyed data quality.

Side elevation of the failed hypermarket setup: one 4K camera ceiling-mounted at 4.5 m with a 120° view over 15 m of shelving, where upper shelves block the lower ones and price tags are unreadable. — One ceiling camera over 15 m of shelving. The upper shelves hide the lower ones.

Failed: refrigerators with internal lighting

A camera above a refrigerated display, bright internal LEDs, glass doors, shooting top-down at 45°.

Glare from glass and LEDs overexposed the image. Reflections covered roughly 60% of the frame, condensation blurred the rest, and once the doors opened, fogging blinded the system for 15-20 minutes. The cause was the physical environment, the glass, strong light and condensation, all predictable and all ignored at design time.

Side elevation of the failed refrigerator setup: a top-down camera at 45° over a glass display with internal LEDs, where glare covers about 60% of the frame and fog blinds it for 15-20 minutes. — Top-down over lit glass. Glare and fog take most of the frame.

Worked: modular system in a premium supermarket

5 MP cameras at 1.8 m, each covering 2-3 m of shelving (one section), 60-70° front-facing field of view, 1.5 m from the shelf.

The system recognized up to 95% of products. Small packaging details were visible, price tags were readable, counting was accurate, and it caught even minor product shifts. The spend on more cameras paid off through data accuracy: a deliberate balance between camera count and coverage quality.

Side elevation of the working modular setup: a 5 MP camera at 1.8 m, 1.5 m from the shelf, with a 60-70° field of view that covers the full shelf. — One 5 MP camera per section, at eye level, 1.5 m from the shelf.

Plan view of the same modular setup from above: one 5 MP camera for each 2-3 m section along the aisle. — The same setup from above: one camera per section along the aisle.

Worked: refrigerators with the right angle

Cameras mounted frontally at mid-shelf height (1.2-1.4 m), 0.8 m from the glass, shooting perpendicular to it (0° to the normal), with a polarizing filter on the lens.

The perpendicular angle and the polarizing filter minimized glare. The short distance offset condensation, and recognition worked in about 90% of cases, recovering within 3-5 minutes of the doors opening. Combining glare reduction, the right angle and close proximity produced stable performance where the top-down setup had failed completely.

Side elevation of the working refrigerator setup: a 5 MP camera mounted frontally at 1.2-1.4 m, 0.8 m from the glass, shooting perpendicular to it with a polarizing filter. — Frontal, perpendicular to the glass, with a polarizing filter. Glare drops away.

The takeaway

Spend the camera budget where it sets the ceiling, resolution and lighting first, then angle and coverage, and test the awkward fixtures (islands, fridges) before rollout. Every point of accuracy you lose at the lens is a point no model will give back. This is what the test phase of a shelf-monitoring project exists to measure, honestly, in your stores, before anyone signs up to scale.