Short answer
Apparel retailers and fulfillment operators hold catalog images and attribute tables for garments: category, fabric, color, closures, size and care. These support recognition, sorting and attribute grounding. They show garments in neat, flat or mannequin poses, not crumpled in a bin, and they say nothing about how a garment moves when grasped or folded.
The buyer's problem
Garments are widely regarded as difficult everyday objects for a robot. They have no fixed shape, they hang, drape, tangle and fold over themselves, and the same item looks different in every pose. Folding, bagging, sorting returns and putting away stock all begin with the same perception step: what is this item, and what is it made of?
Retailers and fulfillment operators answer that question for human shoppers every day. A product page carries a category, a fabric composition, a color name, a closure type and a size chart. This page covers sourcing that attribute data and its images as a training and evaluation asset, and spells out what it leaves out.
What the data can support and what it cannot
It can support:
- Category and subcategory recognition (shirt, trousers, outerwear, socks).
- Attribute grounding: connecting words like "button-front", "ribbed cuff" or "hooded" to visible features.
- Material hints for handling classes, such as stretch knit vs stiff woven, as a prior rather than a measurement.
- Sorting rules based on category, size and care label.
- Language grounding between product text and images.
It cannot support:
- Manipulation. Catalog data does not show how a sleeve folds, how a collar resists or how a knit stretches under a gripper. Folding and sorting control need robot or teleoperation data with contact signals; text and images alone do not train it.
- In-situ appearance. Studio images have even lighting, clean backgrounds and tidy poses. Returns bins and laundry piles do not. See packshot vs in-situ images.
- Physical properties beyond listed attributes. "Cotton" does not tell you thickness, drape or friction. See deformable product properties.
- Condition. New-with-tags listings do not represent worn, stained or damaged items, which matter for secondhand and recommerce item data.
What published product data looks like
The Amazon Berkeley Objects collection pairs catalog images and metadata with artist-made 3D models: 147,702 product listings, 398,212 unique catalog images, and up to 18 metadata attributes per product (Collins et al., 2021). It is a general product set, not an apparel dataset, and it is cited here only as a catalog-format reference: an example of how listing images and attributes can be paired. It says nothing about what any particular retailer holds, and a garment-focused buyer should not treat it as coverage.
Illustrative garment attribute record
The record is Illustrative. All values are synthetic.
| Field | Example | Notes |
|---|---|---|
sku | sku_88213 | Link to retailer master data |
category_path | tops > knitwear > crewneck | Use the supplier's taxonomy plus a mapping |
fabric_composition | listed percentages as text | Label text, not a lab test |
construction | knit / woven / non-woven | Often missing; derive carefully |
closures | zipper, buttons, none | Affects grasp points |
color_primary | "navy" | Names differ from measured color |
fit_and_length | regular, cropped | Affects drape and fold |
care_label_symbols | text codes | Useful for sorting |
images | front, back, detail, on-model | Count and types per SKU |
image_rights | photographer, brand, license status | See below |
Add a column for season or collection_year if you want to test generalization to new styles across time.
Fields that help handling and sorting
Some attributes matter more than others for a robot sorting or folding clothes:
- Size and length category, because a folding routine varies.
- Closures and fasteners, which affect where an item can be grasped.
- Layers and linings, which change thickness.
- Embellishments such as sequins or hard trims.
- Care label text, which is a sorting rule in laundries and reverse-logistics.
- Garment set membership, for matching pairs and sets.
Acceptance criteria you can set
- Attribute completeness. What share of SKUs carry each field? Set your own thresholds.
- Vocabulary control. Are attribute values drawn from a controlled list?
- Image coverage. How many views per SKU, and are details like closures visible?
- Duplicate and variant handling. Colorways of the same style are linked.
- Label accuracy. A spot check by a human reviewer for category and fabric.
- Rights clarity. A rights status per image, not only per account.
Preparation work to expect
Normalizing category trees across suppliers, resolving color names, deduplicating near-identical listings, parsing composition strings, and dropping records where the image was removed or the license lapsed. Treat a catalog extract as raw material for a schema such as the object attribute schema, not as a finished dataset. The wider catalog view is in product catalog data for robot perception.
Questions to ask suppliers
- Who photographed the images: staff, agencies, brands or marketplace sellers?
- What rights apply to each image, and for what uses?
- Which attributes come from the brand and which were entered by retail staff?
- Is fabric composition from the garment label, a supplier sheet or an estimate?
- Do records include discontinued and out-of-season items?
- Are there images from returns processing or warehouse intake as well as catalog photos?
- How are variants such as size and color linked?
- Can the extract be delivered as structured data with stable identifiers?
Permissions and provenance
Apparel imagery carries photographer rights, model releases and brand trademarks. A retailer may not hold the rights to brand-supplied images, and marketplace listings may be seller-owned. Images with identifiable people raise consent questions beyond copyright. Ask who can license each class of image, and see product image rights for the full list of questions. This is not legal advice.
Where this fits
Apparel attributes are a perception and language layer. For contact behavior, a humanoid program still needs robot data. See object recognition data vs manipulation training data.
Discuss your requirements
If you need garment images and attributes for sorting or folding research, describe the categories, attributes, image types and permitted uses to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable catalog data, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. Explore the objects hub, the humanoid training data guide and the catalog category for product catalogs and descriptions.
Discuss your humanoid training-data requirements with SourceX
Sources
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — catalog images, metadata and counts as a format reference.