The hardware question in a shelf-execution project is almost always asked in the wrong order. Someone specifies shelf-edge cameras or a rail-mounted rig, prices a store-wide rollout, and only then asks what the detection task actually needs from an image. By that point the capture spec has been decided by a vendor quote rather than by the shelf-execution classes in scope, and the project’s timeline belongs to a capital cycle instead of to an engineering plan.
Reverse the order. Decide which shelf-execution classes you are detecting — planogram compliance, on-shelf availability, price-tag and promotional checks — derive the minimum image conditions each one needs, then measure the estate you already own against that spec. The output is not a yes/no on procurement. It is a map: which aisles and SKU classes are already servable, and which ones are not.
What does the hardware reuse vs new procurement decision actually mean here?
It means splitting the store, not the project. Reuse and procure are not two competing programme designs; they are two answers that coexist across a single estate. A grocery store with existing loss-prevention cameras will typically find some aisles where the mounted angle and effective resolution are adequate for stock-out detection, others where the same cameras can see presence but not read a price tag, and a handful — deep chilled cabinets, high-facing shelving, glare-heavy end caps — where nothing currently installed will ever meet spec.
The naive path treats the third group as justification for replacing all three. That is the expensive mistake, and it is the one we see stall pilots most often: a subset failure gets generalised into a store-wide requirement, and the whole pipeline waits on approval it did not need.
The minimum capture spec, by shelf-execution class
Each class imposes a different constraint. Effective pixels per facing — not sensor megapixels — is the number that matters, because it accounts for how far the camera sits from the shelf and how much of the frame the fixture occupies.
| Shelf-execution class | Dominant constraint | Typically served by reuse? |
|---|---|---|
| On-shelf availability (gap / empty-facing detection) | Coarse spatial coverage; tolerant of moderate resolution and oblique angle | Often yes — existing overhead and LP cameras frequently suffice |
| Planogram compliance (correct SKU in correct position) | Facing-level discrimination between visually similar SKUs; needs near-perpendicular viewing angle | Sometimes — depends on angle and shelf depth, aisle by aisle |
| Price-tag and label verification (OCR) | High effective pixels on a small target; stable lighting, no glare | Rarely from fixed estate — usually mobile capture or dedicated cameras |
| Promotional and display compliance | Wide-area context, lower per-facing detail, low cadence | Usually yes — often the easiest reuse case |
The most useful thing this table encodes is that the reuse decision is per class and per aisle, not per store. A single answer for the whole estate is a sign the spec was never derived.
Cadence belongs in the spec too. A stock-out on a fast-moving line needs observation frequency measured in minutes to hours, while promotional compliance can be checked daily. Cadence requirements often drive the capture decision harder than resolution does, because a fixed camera gives you cheap frequency while a mobile device gives you sharp images at whatever interval staff happen to walk the aisle. We cover the latency side of that trade-off in how shelf-execution AI catches stock-outs sooner than store-staff rounds.
Testing the estate before writing a procurement request
The test is cheap and it is empirical. Pull sample frames from the existing cameras covering a representative set of fixtures, run them through the candidate detection pipeline, and record where the pipeline’s confidence collapses. Frames from staff mobile devices go through the same pipeline. Nothing about this requires new hardware, and it is exactly the profiling step our capture-and-inference audit performs before any procurement conversation opens.
A practical sequence:
- Enumerate the shelf-execution classes in scope and their cadence requirement per class.
- Sample frames from every distinct camera geometry in the store — not every camera, but one per mounting pattern.
- Measure effective pixels per facing and viewing angle from those frames, rather than from installation drawings, which are frequently stale.
- Run the detection pipeline on the samples and record per-class pass/fail by geometry.
- Project the pass/fail map across the estate by camera geometry to get the reuse share.
- Cost new capture only for the geometries that failed a class you actually need.
Step 5 is where most of the money is saved. Stores in a chain repeat a small number of camera geometries, so a test on one or two stores usually generalises across the format — an observed pattern from retail-CV work rather than a benchmarked figure, and one that breaks down when store formats are genuinely heterogeneous.
Where staff mobile devices work as the capture layer
Mobile capture solves the resolution problem and creates a coverage problem. A modern store-associate handset produces images comfortable for price-tag OCR and SKU-level planogram checks at reading distance, and the device is already deployed, already managed, and already carries a per-frame location context if the capture is bound to a shelf or fixture identifier.
What it does not give you is scheduled observation. Mobile capture is opportunistic — the aisle is measured when someone walks it. That makes it strong for planogram audits and price-tag verification, where a daily or per-shift cadence is acceptable, and weak for high-velocity stock-out detection, where the observation interval is the whole point. Where mobile capture also tends to disappoint is consistency: handheld angle, distance, and motion blur vary far more than a fixed mount, so the model sees a wider input distribution and needs training data that reflects it.
What genuinely justifies new hardware — and how to scope it
New capture is justified when a class you need cannot be served at the required cadence by any existing sensor covering that fixture. That is the whole test. It is not justified by a preference for uniform infrastructure, nor by a vendor’s coverage model, nor by a single failing aisle.
Once justified, scope it to the failing unit. If chilled cabinets fail glare-limited OCR across the format, procure for chilled cabinets. If four end caps carry the promotional SKUs that drive most of the availability loss, instrument those four. The relevant comparison is the per-aisle cost of new capture against the time-to-restock improvement it unlocks on the SKUs in that aisle — a fixture-level business case, not a store-wide rollout budget. Aisles carrying low-velocity, low-margin lines frequently do not clear that bar, and saying so early is more useful than a coverage percentage.
Judging whether the reused estate is actually performing takes the same measurement discipline as the instrumented aisles; that is a separate exercise, treated in how to measure on-shelf availability lift from shelf-execution AI. And the procurement-sequencing failure mode itself — pilots that hand their fate to a budget cycle — is developed further in our retail shelf-execution and on-shelf availability work.
What follows once capture is settled
The capture answer constrains everything downstream, which is why it comes first. Fixed cameras at short cadence push you toward edge inference — a small accelerator per store, models exported through ONNX and served with TensorRT or an equivalent runtime, so that frames are not shipped over a store’s uplink at volume. Mobile-first capture inverts that: image volume is low and bursty, so a central inference service is usually simpler and cheaper, and the network question becomes one of upload reliability rather than sustained bandwidth. Mixed estates need both paths, and the routing rule should be written down before the first camera is bought.
Our broader treatment of shelf-execution systems, including where the computer vision engineering work sits relative to the retail operations layer, gives the fuller picture of how these pieces fit.
If the capture spec is derived before the procurement request is written, the reuse-versus-procure question stops being a budget argument and becomes an engineering measurement. What is your reuse share by shelf-execution class — and do you currently know it aisle by aisle, or store by store?
Frequently Asked Questions
What minimum capture spec does each shelf-execution class actually require?
Availability detection needs coarse coverage and tolerates oblique angles; planogram compliance needs facing-level discrimination and a near-perpendicular view; price-tag OCR needs high effective pixels on a small target plus stable, glare-free lighting; promotional compliance needs wide-area context at low cadence. The governing number is effective pixels per facing, which combines sensor resolution with camera distance and framing, not sensor megapixels alone.
How do you test an existing store camera estate against that spec before committing to procurement?
Sample frames from one camera per distinct mounting geometry, run them through the candidate detection pipeline, and record per-class pass/fail from the results rather than from installation drawings. Because chains repeat a small number of camera geometries, a test on one or two stores usually projects across a format — an observed pattern rather than a benchmarked rate, and one that breaks down where store formats vary widely.
Where do staff mobile devices work as the capture layer, and where do they not?
They work where image detail matters more than observation frequency: price-tag verification and planogram audits at per-shift or daily cadence. They do not work for high-velocity stock-out detection, because capture only happens when someone walks the aisle, and handheld variation in angle, distance and motion blur widens the model’s input distribution.
How do you compare the cost of new capture hardware against the gain it unlocks?
Cost it per aisle or per fixture, not per store, and compare against the time-to-restock improvement on the SKUs that specific fixture carries. Aisles with low-velocity, low-margin lines often fail that test, which is a legitimate result — the goal is a fixture-level business case, not a coverage percentage.
What infrastructure decisions follow once the capture question is settled?
Fixed cameras at short cadence push inference to the edge — a per-store accelerator with models exported via ONNX and served through TensorRT or similar — to avoid shipping frame volume over a store uplink. Mobile-first capture generates low, bursty volume, so central inference is usually simpler, and the network concern becomes upload reliability rather than sustained bandwidth.
Building your monitoring stack without overbuying
Most retailers discover too late that their existing camera infrastructure can’t deliver the image quality AI models need for reliable shelf recognition.
