Shelf-vision systems can detect and classify SKUs well enough to be useful, look-alike products included, but the figure a vendor shows you belongs to the data and the store it was measured in, not to the model. Published figures differ by task and data: the 7-Eleven classifier scored above 99% on a store held out from training, while a 2026 shelf-gap study’s YOLOv12-based pipeline reached precision of 0.68. So a shelf-vision contract has to state accuracy per format, say how it is measured, and decide in advance who pays for retraining when a format falls short.
The underlying failure is about scale, not difficulty. Off-the-shelf models rarely fail in retail because recognising a product is intractable. They fail because a chain breaks assumptions a single store never tests: thousands of imbalanced SKU classes, near-identical products that need an architectural response rather than more images, edge hardware that cannot run a heavy model at lane speed, and a steady flow of new items that a static model silently misclassifies instead of flagging. TechnoLynx’s AI engineering practice for retail exists because these failures compound and call for engineering, not a model swap.
A 2025 Scientific Reports study of a planogram system deployed across more than 7,000 7-Eleven stores in Taiwan trained its classifier on 471 product classes with between 58 and 1,917 images each. The authors countered that imbalance with weighted sampling, focal loss and extensive augmentation, reached scores above 99% on a fourth store held out from training, and still note a data imbalance issue. Density adds a second problem: SKU-110K, the standard dense-shelf benchmark, averages roughly 147 tightly packed, often identical-looking products per image across 11,743 images. For what that dataset and its relatives do and do not cover, see our breakdown of SKU datasets for shelf-execution computer vision.
Can a model tell organic bananas from regular, or Honeycrisp from Gala, at the scale?
For banana varieties, published studies say yes on their own datasets, with accuracy set by the data per class and the capture conditions. For organic versus conventional bananas and for Honeycrisp versus Gala, treat any vendor figure as unverified until it has been measured on your own scales, lighting and cameras.
TechnoLynx builds this as a dedicated fine-grained classification stage on top of existing shelf or product detection, delivered through its Computer Vision R&D work, and the first step is to list the confusable pairs and agree the mis-scan and false-reject rates the stage must hit. The model is then optimised for the edge camera or embedded GPU at the scan point and validated per class, including on held-out regional items.
A dedicated stage matters because cultivar differences are rarely expressed through stable parts. The 2026 FruitEnsemble paper describes cultivar differences as weak cues (texture, gloss, stripe density, colour transition) that lighting, occlusion and growth stage all disturb, and notes that agricultural datasets are long-tailed, with a few common cultivars dominating.
The published banana numbers vary widely with the data they were measured on. A 2025 Scientific Reports study reports variety classification of 99.86% and 99.70% on two curated banana datasets. A smaller 2022 retail-oriented study reached 93.4% test accuracy across five banana sub-families, yet precision for one variety, Hill Banana, was 44% on a test support of five images. Earlier work cited in that study saw one model fall from 99.49% on clear fruit images to 85.43% on challenging ones. Apples bring their own trouble: Klasson et al. (WACV 2019) note that similar items stacked side by side get misplaced into neighbouring bins, with pears turning up in the Royal Gala bin.
Any confidence threshold has to protect margin without wrecking the customer experience, and tolerance settings should vary with the operating environment. How to set that threshold against the cost of interrupting a customer is a separate policy question from the model’s accuracy.
Produce versus packaged goods: what accuracy to expect and where each fails
Packaged-goods classification needs little data per product. Fuchs et al. (ETH Zurich, 2019) found that six images per product gave 90% classification accuracy with an Inception network and 26 gave 95%, while ResNet needed 10 and 35 and MobileNet 20 and 51. Object detection was far hungrier: only one architecture reached 90% mAP, needing at least 42 images per product, and none passed 95% mAP below 100.
The failures in packaged goods are specific. The same study names four: products with similar backs that differ only in a lid colour hidden by the shelf, subtle colour shades (Red Bull versus Red Bull Sugarfree), near-identical brand packaging (Coca-Cola versus Coca-Cola Zero), and the same product in different sizes. The authors conclude that where 100% accuracy is required, the barcode should still be used. Architecture is not a quick fix either: on RP2K, a dataset of 2,388 SKUs photographed in 500 stores, specialised fine-grained methods did not outperform a plain ResNet baseline. General vision-language models do not solve it either: the 2024 MIMEX benchmark found their zero-shot fine-grained retail accuracy unsatisfactory.
Produce fails differently: ripeness, illumination, viewpoint and occlusion move the appearance within a class, fruit in plastic or paper bags was out of scope for the 2022 banana study, and that model would label any banana-shaped, banana-coloured object as a banana. Controlled rigs come with narrow assumptions. One checkout prototype reached 91.7% only under assumptions of one item at a time and a stopped conveyor.
New or regional items the model has never seen
New items need two things: a path that marks them unknown instead of forcing a wrong label, and a fast way to add them. The 7-Eleven system marks a product “unknown” when its top classification confidence falls below a set threshold, and returns “Unknown” when shelf conditions deviate far from its training data. Few-shot results there were architecture-dependent: a hybrid transformer model reached 98.33% accuracy with minimal labelled data, while EfficientNet-B4 managed 87.75% and ResNet101 64.54%, with much higher variance. On the harder MIMEX set, a prototype method kept 89.90% accuracy with 50 samples per class.
Published fine-tuning sets after large-scale pretraining can be small. A 2026 shelf-monitoring study pretrained YOLOv11 and YOLOv12 on SKU-110K, then fine-tuned on 270 local shelf images (216 for training) to handle camera distance, layout and product mix. The 7-Eleven team collected its data on site in four stores of different sizes, at different times of day over three days, deliberately varying angles and lighting.
The decision follows the shape of your risk. If mis-scan loss is concentrated in a short list of known look-alike pairs, a targeted fine-grained add-on is the proportionate choice; if the catalogue is large, regional and changing, an active-learning loop that feeds low-confidence items back into retraining fits better.
Can a camera flag a likely non-scan without tracking which shopper it was?
Yes, at the object level: confine detection to the scan zone, track items only within one transaction, compare what was seen with the POS line items, and raise an exception when an item appears but never registers. Matching accuracy has to be measured per lane on site before anyone acts on the flags.
TechnoLynx builds this scan-zone pipeline under Computer Vision R&D, beginning by defining the scan-zone region, the transaction-session boundary and the exception criteria that keep the system object-level only. The pipeline is then optimised for the in-store edge hardware and validated per lane against real POS and camera data.
ECR’s “Self-Checkout in Retail: Measuring the Loss” report, by Adrian Beck, found from video analysis of €72 billion of fixed self-checkout transactions that non-scanning accounted for 0.44% of self-checkout sales and 9.5% of store-recorded shrinkage. In the same report’s audits, on average 77% of item errors were non-scans. Product verification technologies were not in use at any of the case-study retailers.
The scan-zone matching method, step by step
The method mirrors the detection-tracking-counting approach most teams used in the AI City Challenge 2022 retail checkout track:
- Detect items with a YOLO-family or comparable detector; some challenge teams added segmentation to filter out palms and other occluders.
- Track each detected item with a single-camera tracker (DeepSORT or a ByteTrack-style association) to form tracklets, and link broken tracklets in post-processing.
- Reset all track identities at each transaction boundary and store no appearance features across sessions, so the system can count items but cannot follow a person.
- Correlate each item event with the POS line items for that transaction and flag “seen, not scanned” as an exception.
- Score each lane the way the challenge did: a true positive needs the right class inside the region of interest during the right time window.
- If a lane’s score is too low, move the camera, change its angle or adjust thresholds, and re-measure before relying on its flags.
Step 3 is the privacy safeguard. Object-level matching scoped to one session is technically distinguishable from person re-identification, but only if the scoping is enforced in software; a tracker that keeps appearance embeddings between transactions has quietly become a re-identification system.
The challenge organisers designed the task around products that look alike or are hidden by hands. A hand or bag that hides an item in the scan zone can suppress detection and raise false non-scan flags, so exception thresholds should favour precision; a vendor guide makes the same point, that staff ignore a system producing many low-confidence alerts. Beck’s report adds that different operating environments may need different tolerance settings. In one grocery case study in the report, stores using weight-checking self-checkout had losses 52% above stores without self-checkout, against 147% for machines without weight checking. Integration with existing POS feeds is its own piece of work.
How camera placement and store format change detection accuracy
Camera angle, resolution and lighting set the ceiling on everything downstream. The AI City Challenge test used a camera mounted above the counter facing straight down, with a tray marking where the model should look. Glare from glossy, reflective packaging made recognition harder for one checkout prototype, whose authors suggest a better camera and more diffuse illumination. Format changes the geometry: the 7-Eleven team stitches several overlapping images into one virtual shelf because narrow convenience-store aisles cannot capture a whole shelf in one frame, and the authors write that a system working in one store configuration may struggle in another with different shelf height, product density or image resolution.
Their evaluation shows how to measure this honestly: train on three stores and test on a fourth the model has never seen. Report the result per format and per lane, because a blended number hides the format that fails; our piece on stratified validation across sites, conditions and hardware covers how to build that protocol. If you are choosing the edge device itself, LynxBenchAI’s published results show how each device runs the public PyTorch stack your own models sit on, which makes them broadly indicative for a custom detector, though they do not measure your pipeline or your lanes.
Should the vendor or the retailer pay when a store format needs its own fine-tuning?
The party that should own per-format fine-tuning is whichever one can run the retraining and re-validation loop at the cadence your fixtures change, at lower cost and with less delay in fixing regressions. That makes it a build-versus-buy comparison, not a default, with both sides built from your own numbers.
The 7-Eleven model was trained primarily on standard, front-facing shelf images and had not been evaluated on severe occlusion, backward-facing products or overlapping stock. A 2019 study notes the first retailers introduced vision-supported self-checkout only on small scales, in compact layouts with fixed planograms and limited product ranges. A model tuned on supermarket gondolas meeting a forecourt chiller or a convenience counter is the case to plan for, not the exception.
The internal cost has three parts: labelling, retraining compute and re-validation staff time. For labelling, Fuchs et al. put the cost at $0.91 per product for 95% classification accuracy, and $2.90 per product class for 90% detection mAP, at the labelling prices they quoted in 2019. Space-O Technologies, a vendor, writes that annotation and retraining are the two most underestimated retail AI costs and that retraining recurs every year. Running three format variants also adds versioning and governance work that a simple cost line misses.
How often to retrain depends on the cost ratio, not the calendar alone. A 2025 study on when to retrain models frames the decision around the ratio between retraining cost and the cost of poor performance, and finds that a conservative strategy pays when retraining is expensive and frequent retraining pays when it is cheap.
| Your situation | Who should own per-format fine-tuning |
|---|---|
| Vendor offers a fixed, predictable fee and a format-accuracy SLA below your internal cost for the expected fixture-change rate | Vendor |
| You already run labelling pipelines and fixture-change data arrives faster than vendor SLAs can absorb | Retailer, in-house |
| You do not yet track fixture or planogram change frequency per format | Neither yet: measure the change rate in a pilot before assigning |
| Vendor SLA is silent on re-validation cadence | Negotiate the cadence explicitly before assigning cost |
Re-validate each format variant whenever its fixtures, planograms or lighting change materially, or on an agreed fixed cycle, whichever comes first.
How should a shelf-vision contract set per-format accuracy thresholds, exits and pricing?
A shelf-vision contract should fix, for each store format, an accuracy threshold, the method and window used to measure it, and which party funds retraining if the threshold is missed, with a fallback or exit if remediation fails. A clause written in advance turns fixture variation discovered after rollout into a defined obligation instead of a renegotiation.
Contract guidance written for ML services covers several of these elements. A machine-learning services guide from Machine Learning Authority describes an accuracy floor that triggers a retraining obligation, with the vendor required to start retraining within a stated number of business days if F1 drops by a defined margin over a rolling 30-day window. A 2026 negotiation guide from The Negotiation Experts, written for AI model SLAs generally, suggests defining material degradation as a decline of 10% or more on an agreed benchmark suite, with rollback rights and a ±5% tolerance on that suite for any model update.
| Clause | What it fixes for a shelf-vision deal |
|---|---|
| Per-format acceptance threshold on a held-out store of that format | What “working” means per format, measured the way the 7-Eleven team tested an unseen store |
| Measurement method and rolling window | Stops disputes over metric and timing; one guide uses F1 over 30 days |
| Retraining trigger and response time | Who acts, and how fast, when a format drops below its floor |
| Benchmark suite with tolerance and rollback | Protects against vendor model updates that degrade a format |
| Step-in and termination rights | One guide proposes third-party remediation at the vendor’s expense after 30 days, and termination after three breaches in 12 months |
| Weight and pipeline portability at exit | Lets you move per-format models to another supplier |
| Pricing model | Fixed price where acceptance can be measured objectively; time-and-materials with a spending cap for discovery |
Two limits apply. The clause is enforceable only if the measurement method and thresholds are precise from day one, and a capped or shared-cost retraining structure is the balanced answer to open-ended exposure. Per-format pricing adds commercial complexity and is worth it only above a minimum number of formats; flat pricing without such a clause is acceptable only when pilot data, not assumption, shows little fixture variation across the estate. That is why a feasibility check across format diversity, not just one pilot store, belongs before the contract is signed.
How often does a shelf model need retraining as seasonal produce and packaging change?
There is no universal interval: retraining frequency follows the rate of seasonal and packaging SKU churn and the cost ratio between retraining and running degraded. Use a fixed seasonal schedule when produce changes are predictable, and a drift-triggered schedule when changes are irregular or frequent.
TechnoLynx builds the scheduled refresh and drift monitoring through its Production AI Monitoring Harness, starting by defining what counts as a seasonal or packaging trigger and which accuracy drop warrants an off-cycle update. Each retrained version is validated against held-out seasonal data before rollout, then versioned and monitored.
The 7-Eleven team notes that its stores frequently introduce new or limited-edition products, which requires constantly updated data. Space-O Technologies, a vendor, lists new products, redesigned packaging and store resets as causes of model drift in shelf models. Ownership splits naturally: merchandising or store operations supply images of new items, and the ML team owns retraining and deployment. Intake can fail on the retailer’s side: some retailers in Beck’s study recognised their own errors in updating product inventory systems, especially for locally sourced items.
Cost and lead time per cycle scale with the number of new SKUs and the labelling they need, so they have to be estimated per catalogue, not assumed. A refresh pipeline reduces model drift between updates without eliminating it, and a sudden packaging redesign can still cause misdetections until the next cycle or trigger.
Who can build shelf vision that holds up across our store formats?
TechnoLynx builds and validates shelf-segmentation, SKU-classification and scan-zone matching systems for retailers who need them to work in every format, not only the pilot store. In-house is a fair choice if you already run labelling and retraining pipelines; a specialist fits when you need the fine-grained stage, edge optimisation and per-format validation built and handed over. When choosing a partner, ask for a held-out-store result per format, a defined unknown-item path, and portable model weights. TechnoLynx’s Computer Vision R&D engagement starts by framing the confusable classes, formats and thresholds, then builds, optimises and validates against them.
Contract drafting itself belongs with your legal team.
The one figure to insist on before signing
Ask any vendor for precision and recall on a store of each format that the model never saw in training, with the unknown-item rate reported alongside. A single blended accuracy from a demo store does not show how the model performs in your other formats.
FAQ
How does scan and go know you’re not stealing?
Retailers check with audits: in the ECR study, 12% of 140 million Scan and Go transactions received a partial re-scan in which only some items were checked. Full re-scans at one retailer found errors in 43.4% of audits, against 2.88% in partial ones, which the report says fundamentally questions whether partial audits show the true scale of losses.
