Every CPG brand manages execution at stores it doesn't control. The field rep's visit is the only window to verify what the shelf looks like—whether the planogram held after the last reset, whether the promotional display is still in place, whether the price tag reflects this week's campaign or last week's.
When that verification method is a rep doing visual checks across hundreds of SKU positions during a 15-minute store visit, the accuracy floor sits at 60–70% for position-level deviations.
The facing count that dropped from 2 to 4, the SKU that drifted one position left of its planogram slot, the price tag from the previous promotion still in place—these are the gaps that consistently slip through and accumulate into material revenue losses between visits.
Image recognition changes what happens during that visit.
A rep photographs the shelf section, computer vision reads every visible SKU, maps each one to its planogram position, reads price tag values, and returns a ranked gap list to the rep's phone within 90 seconds. The rep corrects what's fixable before leaving the aisle. HQ sees the before-and-after from the same visit in real time.
Both systems use cameras, but that's where the similarity ends.
Surveillance captures movement and presence. A CCTV camera above the beverage aisle registers that a product was picked up, that someone walked through the section, or that a shelf area was accessed. The output is behavioral data—what happened, when, and roughly where.
Image recognition reads identity and position. When a field rep photographs the same beverage aisle, IR identifies each individual SKU by brand, variant, and size. It maps every product to its exact shelf position, counts facings per SKU, reads price tag values, and compares every position against the approved planogram. The output is execution data—what is currently on the shelf, down to the SKU level, and whether it matches the plan.
A CCTV system can tell a loss prevention team that a product was removed from the shelf, but it can’t tell a category manager that the hero SKU dropped from four facings to two, sits one position left of its planogram slot, and has a price tag from last week's promotion still in place.
That distinction—behavioral visibility vs. execution accuracy—is what makes image recognition a category management tool rather than a security tool.
ERP systems have inventory data, and POS systems have sales data, but neither tells a category manager whether the units the ERP says are in the store are actually on the shelf, in the right positions, with accurate pricing, and with the correct promotional materials installed.
The visual truth gap is the distance between what the data says and what a shopper actually encounters.
A beverage brand's ERP shows 240 units of the hero SKU in a store. The POS shows 40 units sold last week. Neither system records that 200 of those 240 units are in the backroom, that the four-facing eye-level position on the planogram has been collapsed to one facing after a reset, or that the promotional price tag reflects last month's campaign.
That gap is where execution investment leaks. Planograms that were approved at HQ but didn't survive the reset. Promotions that were funded but ended up in low-traffic bay positions instead of the contracted front-of-store endcap. Price tags that weren't updated when a new campaign started.
Image recognition technology closes the gap by reading the actual shelf state during the visit and comparing it against the plan. It produces a realogram—a record of what the shelf actually looks like—and surfaces the difference between that and the planogram in the same moment the rep is standing in front of the shelf.
The process runs automatically from the moment a rep captures a shelf photo. Five stages happen sequentially, typically completing within 90 seconds:
A field rep photographs the shelf section in overlapping frames—three to four photos cover a standard 12-foot gondola run. The app guides the rep through the capture sequence, flagging whether frames have sufficient overlap and lighting quality before submission.
Computer vision scans the assembled image and locates every product.
The system identifies individual items by placing a bounding box around each one, separating products from shelf edges, labels, price rails, and background. Every visible product gets a set of position coordinates within the image before recognition begins.
Each detected product is matched against the reference database using multiple simultaneous signals. This step—fine-grained visual categorization—is what separates enterprise IR from basic object detection.
The AI reads packaging shape, label text via optical character recognition, color palette, dimensional ratios, and brand mark positioning all at once. No single signal is reliable in isolation. Combined, they achieve the 95%+ accuracy benchmark at the SKU variant level that makes enterprise deployment viable.
This is how IR distinguishes between a Low-Sugar and Original variant of the same cereal brand when the packaging is 95% identical. The label text reads differently, the color treatment uses a slightly different palette, and the variant callout sits in a different position on the front panel. The AI reads all three simultaneously.
Each identified SKU is mapped against the planogram on file for that specific store.
The system flags position deviations, missing facings, wrong products in allocated slots, price tag discrepancies, and POSM presence or absence. Deviations are ranked by commercial priority—a missing facing on the top-selling SKU in the category weighs more than a minor POSM orientation issue.
Results reach the rep's phone within 90 seconds of the final photo.
The gap list shows exactly which SKU, which position, and what the deviation is. The rep works through the highest-priority corrections before moving to the next section. Corrections are photo-documented, and HQ sees the before-and-after shelf state from the same visit in real time.
The hardest recognition problem in retail isn't identifying whether Coca-Cola is on the shelf but identifying whether the facing is Coca-Cola Original 330ml or Coca-Cola Zero Sugar 330ml when both use a red can, the same brand logo, and differ only in the label text and a color accent on the ring pull.
Fine-grained visual categorization (FGVC) is the technical capability that handles this. It works by reading multiple signals simultaneously rather than relying on any single visual attribute:
The combination of these signals is what makes 95%+ accuracy at the variant level achievable. A system relying only on color would confuse Coca-Cola Zero with Diet Pepsi. A system relying only on shape would confuse all 330ml cans. Reading all four signals simultaneously produces the variant-level specificity that matters for planogram compliance.
Where the AI model runs determines whether the tool works reliably in real field conditions.
In a cloud-processing setup, the rep photographs the shelf, the images are uploaded to a server, the recognition runs in the cloud, and results are returned to the device.
The analysis is as good as the connectivity. In a rural grocery with poor signal, a basement-level pharmacy, or a convenience store with thick concrete walls, the round trip slows dramatically or fails entirely. The rep waits, gets frustrated, and stops using the tool.
In an edge-processing setup, the recognition model runs locally on the device.
The rep's phone processes the image. No connectivity is required for the analysis step—results arrive in seconds regardless of network conditions. Data syncs to HQ servers when connectivity resumes, but the rep gets the gap list in the store whether or not they have a signal.
For a field rep covering 10 stores in a day across a mixed-connectivity territory, edge processing is the difference between a tool that works everywhere and a tool that works when conditions cooperate. Enterprise IR platforms built for CPG field execution run edge processing as the default.
A common concern: if the model was trained on existing products, what happens when a brand launches a limited edition holiday pack, reformulates a product, or makes a packaging change?
Mature IR platforms handle new SKU introductions through automated model update cycles. When a new product needs to be added to the recognition library, the brand submits several high-resolution front-facing product images. The platform triggers a retraining cycle that incorporates the new SKU into the model. Most platforms run this process on a 2–3 week cadence, meaning a new product launched today is typically recognizable within two to three weeks of image submission.
Vision Group’s Store360's pre-trained library of over 1.3 million SKUs across CPG categories already includes the majority of mainstream branded products before a client deploys the platform. For most CPG brands, the products they sell in their top accounts are already in the library. New launches, limited editions, and packaging refreshes are handled through an automated intake process rather than a manual configuration project, which is a key reason most clients go live in under 30 days.
Image recognition technology has existed in retail for a decade. Enterprise adoption at scale is a recent development. Three specific shifts explain why the pilots that stalled in 2019 and 2021 are turning into full network deployments in 2026.
Early commercial IR deployments in 2018–2020 achieved ~80% accuracy under real store conditions. At that rate, one in five reads contains an error—too high for operational decisions.
By 2024–2025, leading platforms reached 90–95%+ accuracy under standard field conditions. At 95%, one in twenty reads contains an error, and human-in-the-loop validation handles those exceptions.
The improvement from 80% to 95% is the difference between a technology that requires constant manual correction and one that earns an operator's trust.
Advances in mobile chip performance—specifically Apple's A-series and Qualcomm's Snapdragon processors—made it practical to run complex computer vision models on a standard field rep's phone without requiring cloud connectivity. This eliminated the signal dependency that made early IR deployments unreliable in mixed-connectivity field conditions.
A rep in a rural store with no signal gets the same 90-second gap list as a rep in a metropolitan flagship store.
The 8–16 week product data collection and model training process that killed early CPG adoption programs has been replaced by pre-trained libraries that already recognize most mainstream branded products.
What was a multi-month IT project before any useful data appeared is now a 30-day deployment for most enterprise CPG clients. That timeline change alone moved IR from a strategic initiative requiring board approval to a tactical deployment a field execution director can own.
Before IR, retail execution ran on a simple and structurally limited loop.
A rep visited a store, checked the shelf by eye, recorded findings on a form or a mobile app, and submitted a report. A manager reviewed the report—usually 24–72 hours later—and decisions were made based on what the shelf looked like at one point in time, recorded at whatever accuracy the rep achieved under route pressure.
IR breaks that loop at two points.
First, it replaces the rep's subjective visual check with computer vision that reads every SKU position at 95%+ accuracy.
Second, it delivers findings to the rep before they leave the store rather than to a manager's dashboard hours later.
Those two changes—accuracy and timing—compound into fundamentally different execution outcomes.
A category manager running a manual audit program reviews monthly compliance averages.
A category manager running an IR program reviews store-level execution data from every visit—which three stores in the Northeast have had back-to-back planogram failures on the cola section, which rep's route includes them, and how many consecutive visits the deviation has persisted.
The first data set supports quarterly reviews. The second supports weekly corrections.
An execution failure caught during the visit costs the brand hours of sub-optimal shelf time.
The same failure caught in a post-visit report and corrected on a follow-up trip costs the brand days—or weeks if the next scheduled visit is two weeks out.
For a hero SKU doing $50,000 in weekly sales at a high-traffic grocery account, the correction timing difference has a specific dollar value. IR makes that correction happen on the same trip the failure was detected.
A CPG brand that pays for four eye-level facings in the cola section at a major retailer has historically had to trust that the planogram was executed correctly.
With IR, every store visit produces photographic documentation of whether those four facings are in place, at the correct height, with accurate pricing, and with the contracted display materials installed.
That documentation supports both compliance verification with the retailer and trade negotiation conversations backed by visit-level data rather than quarterly estimates.
Trade promotions represent a significant portion of CPG marketing budgets—and most brands have limited visibility into whether those promotions are actually executing in stores during the campaign window.
IR tracks promotional display presence, POSM installation, and promotional pricing compliance on every store visit throughout the campaign.
A trade marketing director can see mid-campaign which stores are executing correctly and which ones have displays in the wrong position or missing the price-drop wobbler—while there's still time to act.
The value of IR extends beyond the individual store visit.
Visit-level execution data aggregated across a network reveals which planogram configurations hold consistently and which break down—by store format, by banner, by region. When that data connects to planogram management tools and assortment optimization platforms, it closes the loop between what gets planned and what actually executes.
A planogram built with knowledge of which layouts survive real-world execution is a better planogram than one built without it.
IR is a mature technology, but understanding where accuracy degrades prevents unrealistic deployment expectations.
None of these limitations prevent enterprise deployment. They're operational parameters to build into program design—calibrating visit frequency, managing planogram file coverage, and setting accurate expectations about what the accuracy benchmark means in production.
A common question at the enterprise procurement stage, particularly for brands operating in GDPR jurisdictions: what happens to images of customers and store associates that appear in shelf photos?
IR platforms designed for CPG field execution process shelf images, not customer images.
When a customer or store associate appears in a shelf photo, mature platforms strip identifiable visual information—faces, identifiable clothing, credit cards—in real time at the device level before any image data is stored or transmitted. PII never enters the data pipeline.
The audit record that reaches HQ contains SKU positions, compliance scores, and shelf state data—not images of people. Ask any IR vendor to describe their PII handling process specifically before a procurement conversation, and confirm that stripping happens on-device rather than after cloud upload.
Store360 is Vision Group's image recognition platform for CPG brand field execution. It's built specifically around the use case this article describes: a field rep visiting stores the brand doesn't own, where the visit is the only window to detect and correct execution failures.
The core workflow:
A rep photographs the shelf section. Store360's computer vision reads every visible SKU against the approved planogram for that specific store and returns a ranked gap list to the rep's phone within 90 seconds—on-device, regardless of connectivity. The rep corrects what's fixable during the visit. HQ sees the before-and-after shelf state in real time.
Pre-trained library of 1.3M+ SKUs. Most CPG brands don't need to supply product data before deployment. Products are already in the library. Most clients go live in under 30 days.
No planogram required. Store360 benchmarks against category norms and competitor positions when a planogram file isn't available, so every store in the network produces audit data on every visit—not just stores with current planogram coverage.
Edge processing. Recognition runs on-device. Results arrive in 90 seconds whether the rep has a signal or not. Data syncs to HQ when connectivity resumes.
Connected to the full execution workflow. Store360 connects directly to EZPOG for planogram management and Curate for assortment simulation. Execution data from every visit feeds back into category planning decisions rather than sitting in a standalone compliance dashboard.
On-shelf availability and near-out-of-stocks · Planogram compliance at the SKU and position level · Price tag accuracy including promotional pricing · Promotional display and POSM presence · Share of shelf and competitor positions—all five from the same shelf photos, in the same visit, in under three minutes per bay section.
Results from client deployments:
L'Oréal at Walmart: $50,000+ in replenishment orders across 10 stores in two weeks, moving from audit data that was 2–4 weeks old to live shelf visibility during each visit.
"What takes a couple of minutes now used to take 15–20 minutes. Rep Insights lets our people sell directly in store with data in under 60 seconds."—Michael, Nestlé
General results: 22% fewer out-of-stocks and 600,000+ field hours saved annually across Vision Group client deployments. Live in 55+ countries, running on 1.3M+ pre-trained SKUs.
Store360 is live in 55+ countries, runs on the device a field rep already carries, and most clients go live in under 30 days. No new hardware or retailer permission required.
→ Book a 20-minute walkthrough here.
Image recognition in retail is a computer vision technology that analyzes photos of physical store shelves and converts them into structured, SKU-level execution data. When a field rep photographs a shelf section, the system identifies every visible product by its packaging characteristics, maps each one to its planogram position, counts facings, reads price tags, and compares the full picture against the approved execution standard. The output is a gap list delivered to the rep's phone within 90 seconds—showing exactly what's off, in what position, ranked by commercial priority.
A regular shelf photo records what the shelf looks like—it's documentation. Image recognition analyzes what the photo contains—it's data extraction. The difference is what happens after capture. A regular photo sits in a folder until a manager reviews it. An IR-processed photo produces a compliance score, a deviation list, and a prioritized correction task within 90 seconds. The rep acts on findings before leaving the aisle rather than after a manager reviews the photo days later.
Enterprise IR platforms operating under standard field conditions achieve 90–95%+ accuracy at the SKU level. Accuracy is primarily determined by the quality and breadth of the reference product database the model matches against, and by how diverse the training data was across real-world store conditions. At 95% accuracy, one in twenty reads contains an error—which is why mature platforms include human-in-the-loop validation for low-confidence reads. For comparison, manual audits under real field conditions achieve 60–70% accuracy on position-level deviations.
Fine-grained visual categorization (FGVC) is the IR capability that distinguishes between near-identical products—different variants of the same brand that look 90–95% visually similar. It works by reading multiple signals simultaneously: label text via OCR, dimensional ratios between packaging variants, color palette differences, and brand sub-mark positioning. No single signal is reliable in isolation—combined, they achieve SKU variant-level accuracy that makes planogram compliance checking viable for categories with many similar-looking products.
Computer vision is the broader technology—the capability of machines to interpret visual information from images and video. Image recognition is a specific application of computer vision focused on identifying objects and their attributes. In retail, the two terms are often used interchangeably. Technically, computer vision covers a wider range of applications including movement detection, queue analysis, and loss prevention monitoring, while image recognition specifically refers to identifying what a product is, where it is, and what its attributes are.
A field rep photographs a shelf section using a mobile app. The app captures overlapping frames that together cover the full shelf run. Computer vision processes the assembled image in five stages: object detection locates every visible product, SKU recognition identifies each one by its visual characteristics, compliance comparison maps each SKU against the planogram, gap detection flags deviations by type and commercial priority, and gap list delivery sends the prioritized correction task list to the rep's phone within 90 seconds.
Edge processing means the image recognition model runs on the device itself—the field rep's phone or tablet—rather than requiring images to be sent to a cloud server for analysis. The practical advantage is that results arrive in seconds regardless of network connectivity. A rep in a rural store with poor signal gets the same 90-second gap list as a rep in a well-connected urban store. Platforms that require cloud connectivity for recognition fail in the 20–30% of store visits where connectivity is poor or inconsistent.
Most IR platforms require an official planogram to generate compliance data. Without one, they return no useful output for that store. Store360 is an exception—it benchmarks shelf presence against category norms and competitor positions even without a planogram file. Every store in the network generates audit data on every visit, not just the stores with current planogram coverage. For brands where planogram files are incomplete or outdated across part of their network, this difference determines whether 60% or 100% of stores get audited.
It depends on whether the vendor requires the brand to supply product data. Tools that require client-supplied images and UPC data for model training take 8–16 weeks from contract to first useful data. Platforms with pre-trained libraries—Store360 has 1.3M+ SKUs pre-trained—deploy in under 30 days. This is the single most important deployment question to ask any IR vendor before signing. The marketing timeline and the actual production timeline are often different.
A realogram is a digital record of what a shelf actually looks like at the time of a store visit—the actual shelf state, as opposed to the planogram, which represents the intended shelf state. Image recognition produces a realogram automatically from shelf photos. The comparison between the realogram and the planogram is what generates the compliance score and the deviation list. The term 'realogram' is used specifically to distinguish the actual-state record from the planned-state document.
Mature IR platforms handle new SKU introductions through automated model update cycles. The brand submits several high-resolution front-facing product images for the new SKU. The platform triggers a retraining cycle that incorporates the new product into the recognition library. Most platforms run retraining on a 2–3 week cadence. Platforms with pre-trained libraries that cover mainstream branded products typically include new national launches within weeks of market introduction without requiring manual configuration.
The visual truth gap is the distance between what digital systems report and what actually exists on the physical shelf. An ERP shows inventory levels. A POS shows units sold. Neither confirms whether available units are on the shelf in the right positions, with the correct pricing, and with promotional materials installed. Image recognition closes the visual truth gap by reading the actual shelf state during the visit and comparing it against the plan—producing a verified record of execution rather than an inference from transaction data.
Image recognition detects two types of availability failures that manual audits consistently miss. The first is full out-of-stocks—positions with no product. The second is near-out-of-stocks—positions where facing count has dropped below the minimum threshold, typically to one or two units pushed to the back, which appear stocked to a rep doing a visual pass but effectively invisible to a shopper. By catching both types during the visit, IR enables correction before the position goes fully empty and the sale is lost.
Yes. Enterprise IR platforms read price tags during the same shelf photo that captures planogram compliance and facing counts. The system identifies whether a price tag is present, reads the price shown, and compares it against the expected promotional or everyday price for that SKU at that store. Missing price tags, leftover tags from previous promotions, and pricing deviations are all flagged in the same gap list alongside availability and positioning failures.
Heavy occlusion—customers or shopping carts blocking part of the shelf—narrows the read to what's visible. Severely damaged or missing labels reduce recognition confidence for affected products. Double-stacked products behind the front-facing unit are excluded from share-of-shelf calculations. IR reads the shelf at the moment of capture—it doesn't detect what changed between visits. Most tools require a planogram on file to generate compliance data. None of these prevent enterprise deployment, but they're operational parameters to build into program design.
Every CPG brand has planograms, trade agreements, and promotional campaigns. What most brands lack is reliable, fast verification that those plans are showing up on the shelf—at the SKU level, on every visit, while there's still time to fix what's wrong.
Image recognition provides that verification. Not as a reporting layer that tells a category manager what went wrong three weeks ago, but as a live execution tool that tells a field rep what to fix before they leave the aisle.
The technology has matured to the point where deployment is measured in weeks rather than months, accuracy is high enough for operational decisions, and the data connects to the planning tools where category strategy gets built. The gap between what the shelf is supposed to look like and what it actually looks like is now a solvable problem—not a quarterly average.
→ Book a walkthrough of Vision Group's Store360 here.