Short answer
To judge packaging-variant coverage, check whether the data links each image to a specific SKU and pack version, shows redesigns, seasonal editions and size variants of the same product, and includes images from the shelf or workplace as well as studio shots. Pass a dataset only if it lets you split by pack version and test on a redesign the model has not seen.
The buyer's problem
A recognition model trained on this quarter's packaging can fail on next quarter's. Brands change artwork, add promotional banners, switch bottle shapes, print seasonal editions and run limited flavors. Retailers also see regional packs and multipacks that share a name but not an appearance. For a robot that sorts, restocks or kits products, appearance drift is a steady source of errors rather than a rare event.
This checklist helps you evaluate an offer of product images for variant coverage. For the catalog-refresh process that keeps data current over time, see SKU churn and catalog refresh. For the choice between recognizing instances and categories, see instance versus category recognition.
What published product datasets tell you to look for
Products-10K contains 10,000 SKU-level products frequently bought on JD.com, with both in-shop and customer images and a hierarchical label graph that spans fashion, 3C, food, healthcare and household goods (Bai et al., 2020). Two design choices are worth copying into your requirements: SKU-level labels, so near-identical products are not merged, and a mix of seller and customer imagery, so the domain gap is present in the data.
The RPC retail checkout dataset pairs single-product images under controlled conditions with multi-product checkout images, defines 200 product categories grouped into 17 meta categories, and provides multiple annotation levels across up to 83,000 images (Wei et al., 2019). The pairing of clean exemplars with crowded scenes is the structure to ask suppliers about, because the exemplar set is what you can fit to a catalog and the crowded set is what the robot will see.
Neither dataset documents how packaging changed over time, as far as their published descriptions show. Temporal variant coverage is therefore a property you should require explicitly, not assume from dataset size.
Kinds of appearance change to separate
Lumping every change together hides which ones hurt. A useful typology has four levels. Cosmetic drift alters artwork, colors or callouts while the pack shape stays the same. Structural change alters the container: a new bottle neck, a different carton fold, a switch from pouch to tub. Commercial variants include promotional overprints, multipacks and regional editions. Substitution replaces one product with a look-alike from another brand. Ask the supplier to tag each pair by level, because a model that survives cosmetic drift can still fail on structural change, and the handling consequences differ too.
Illustrative variant-coverage checklist
| # | Check | Pass | Fail | Evidence to request |
|---|---|---|---|---|
| 1 | SKU-level labels | Each image tied to a GTIN or internal SKU | Labels are brand or category only | Label schema, sample rows |
| 2 | Pack-version field | Revision or date range recorded per image | No version information | Version table, capture dates |
| 3 | Redesign pairs | Old and new packaging of the same product present | Only one appearance per product | List of products with 2+ versions |
| 4 | Seasonal and promotional editions | Flagged and linked to the base SKU | Mixed in as separate products | Edition flag, linkage table |
| 5 | Size and multipack variants | Distinct SKUs, labeled as such | Collapsed into one class | Count per base product |
| 6 | Regional and language variants | Label language noted | Not recorded | Market field |
| 7 | Wear and damage | Dents, scuffs, price stickers present | Only pristine units | Damage tags, sample images |
| 8 | Domain mix | Studio, shelf and handling images | Single domain | Image source breakdown |
Score by counting hard fails first. A dataset with no pack-version field cannot support a drift evaluation, whatever its size.
Evaluation design you can ask for
Define a temporal hold-out: train on earlier pack versions and test on the newest, for the same SKUs. Define a novel-variant hold-out: withhold an entire redesign group. Report accuracy separately for unchanged and changed packs. If a supplier cannot supply version dates, you can still hold out by capture site or by image source, but that is a weaker test and should be labeled as such. The novel object generalization evaluation page covers hold-out design more broadly, and the grocery and retail product recognition page covers the retail setting.
Building a variant test set from what you receive
Even a supplier with version metadata may not deliver ready-made splits. Group images by SKU, order the pack versions by date, and label each SKU as unchanged, cosmetic, structural or substituted. Reserve every image of the newest version of a sample of SKUs for test, and keep the matching older versions in training. Then add a second test group of SKUs for which no version appears in training at all. The gap between the two groups shows how much of your accuracy comes from having seen the product before, which matters for catalogs with frequent launches.
Red flags
- Image counts are quoted without SKU counts or version counts.
- "Variants" means color swatches of the same artwork, not packaging redesigns.
- Images were scraped from listings with no capture dates, so versions cannot be ordered.
- Duplicates or near-duplicates across splits, which inflate drift accuracy.
- Product master data and images disagree on size or pack count; run a product master data quality audit sample before purchase.
- Kits and bundles are labeled as single products; see kits and multi-part products.
- Handling stickers, shipping marks and price tags are removed in cleaning, though a robot will see them; see handling marks on packaging.
What this data cannot support
Appearance data identifies products. It does not tell a robot how a particular pack handles: whether a redesigned bottle is more slippery, whether a new carton is stiffer, or how a multipack shifts when lifted. Those need physical-property or interaction data. Appearance coverage also reflects what a supplier sold in a period, so long-tail and regional items are often missing.
Permissions and provenance
Packaging artwork carries brand trademarks and design rights, and images carry photographer rights. Ask who can license training use of both, whether brand owners have agreed, and whether the images came from a retailer's own photography or from third-party uploads. See trademarks in training images. This is not legal advice.
Questions to ask suppliers
- How many distinct SKUs and how many pack versions per SKU are included?
- Is the pack version recorded by date range, revision code or inferred?
- Can you identify redesign pairs, seasonal editions and discontinued items?
- What share of images is studio, shelf and customer-captured?
- Who photographed the images, and who holds the rights?
- Are GTINs present, and have they been checked against product master data?
- How were duplicates handled across the data?
- Will you provide the label guidelines and any inter-annotator agreement figures?
Discuss your requirements
If you know the product families and how often their packaging changes, tell SourceX what your evaluation set needs to cover. SourceX reviews the requirement, follows up about potential US sources, and checks the data and the supplier's licensing permissions before a license is agreed. A request does not guarantee a match. For wider context, use the humanoid training data guide and the product and object data hub, and see the site's entry on product catalogs and descriptions.
Tell SourceX what your evaluation set needs to cover
Sources
- Products-10K: A Large-scale Product Recognition Dataset — Products-10K: SKU-level products with in-shop and customer images and a hierarchical label graph.
- RPC: A Large-Scale Retail Product Checkout Dataset — RPC: single-product images under controlled conditions plus multi-product checkout images, 200 categories in 17 meta categories.