Short answer
Resale and recommerce companies may hold photos and attributes of used goods with condition grades, wear descriptions and defect notes, which adds variety that new-product catalogs lack. The data can support recognition under wear, condition assessment and sorting evaluation. It carries consumer-data risks, uneven grading, and no handling or force information, so it cannot train manipulation.
Why used goods are worth a separate look
(New-product catalog data is covered in product catalog data for robot perception.) Many product image sets show items as they leave the factory: clean, uniform, photographed against plain backgrounds. Robots working in returns processing, donation sorting, repair intake or secondhand retail meet something else. The same model of jacket arrives creased, stained or missing a button; a handbag has scuffed corners; an appliance has a cracked housing. A recognition model trained only on pristine catalog photos has not seen those states.
Resale and recommerce operations routinely handle that variety, and they often produce images and condition notes because buyers need to know what they are getting. That makes them a distinct supplier type from manufacturers and retailers, whose product data is covered in retailer vs manufacturer object data. It also overlaps with returned-goods work; for grading returned items in an inspection context, see returned item condition grading.
What the data may contain
Holdings differ by company, and a buyer should not assume any particular set exists. Typical material in this supplier type may include:
- Listing photos taken by staff or by consumers, from several angles.
- A condition grade assigned during intake, using the company's own scale.
- Free-text descriptions of wear, damage, missing parts or modifications.
- Attributes such as category, brand, size, color and material.
- Intake or refurbishment photos and notes, where the company repairs or cleans goods.
- Outcome data: sold, returned, recycled, discarded.
Catalog-style datasets show what structured product data can look like. The Amazon Berkeley Objects dataset pairs catalog images and metadata with artist-made 3D models: 147,702 product listings, 398,212 catalog images, up to 18 metadata attributes per product, and 360-degree turntable views for 8,222 products (Collins et al., 2021). It covers catalog products, not used goods, so it is only a capture, attribute and format reference here. It does not tell you what a resale company's data contains or how well it is graded.
What used-goods data can support
| Use | What the data contributes | What to check |
|---|---|---|
| Recognition under wear | Same product in many conditions | Whether product identity is labeled, not only category |
| Condition grading models | Images paired with human grades | Grade scale, grader consistency, grade definitions |
| Defect and damage detection | Notes and images of stains, tears, cracks | Whether defects are localized or only described in text |
| Sorting and routing evaluation | Destination outcome per item | Whether outcome is recorded and by whom |
| Long-tail variety | Odd items, mixed brands, no SKU | Coverage by category versus a few popular categories |
The long-tail benefit connects to long-tail object coverage, and wear or damage states relate to object state recognition.
Where the data is uneven
Condition grades are assigned by people using company-specific scales, so "good" at one company may be "fair" at another. Different graders within one company may disagree. Photos follow whatever listing practice the company set: some use a standard backdrop and angles, others accept consumer snapshots with clutter. Descriptions may be marketing-flavored, since sellers want to sell. Free-text notes may mention the former owner, an event or an address.
A buyer should plan to standardize grades if more than one supplier is involved and to measure agreement on a sample. The principle for checking grader consistency is in inter-inspector agreement. Remember that photos taken for a listing show what a seller chose to show, not a systematic inspection; see packshot vs in-situ images.
Illustrative package for condition-graded items
This is an Illustrative package with synthetic values. A buyer would adapt fields to their architecture and the supplier's actual records.
# Illustrative only - not an available dataset
resale_item_package_v0:
item_id: rs_0030421
category_path: apparel > outerwear # supplier taxonomy, mapped later
attributes: {brand_id: brand_031, size: M, color: green, material: cotton}
images:
- {view: front, background: plain, resolution: 2000x3000}
- {view: label_close_up}
- {view: defect_close_up, defect_ref: d1}
condition:
grade_supplier_scale: B # scale definition supplied separately
grader_id: grader_A12 # pseudonymous
defects: [{id: d1, type: stain, location: left_sleeve}]
text_note: free text, scrubbed of personal details
outcome: resold | recycled | donated
excluded: seller identity, buyer identity, addresses, payment data
The grader_id field makes agreement checks possible. The excluded line is not decoration: resale records may link items to consumers.
Consumer data and permissions
Resale data originates with consumers who sold, returned or donated items, and with customers who bought them. Listing photos may include faces in mirrors, rooms, addresses on parcels or labels with names. Listing descriptions may carry personal stories. Photos supplied by consumers raise questions about who owns the image and what the seller agreed to.
Ask what terms consumers accepted when they submitted items or photos, whether those terms contemplate licensing to a third party for AI training, and who the company can grant rights on behalf of. Removing names, blurring faces and deleting addresses reduces exposure but does not settle whether the company may license the material; it remains a question about the original terms. Keep any evaluation of item condition separate from any profiling of sellers or buyers; a condition-grading dataset has no need of customer histories. This is not legal advice. For wider guidance on identifiable people, see worker consent and identifiable people as a model for the diligence approach, and the catalog for images and inspection photos for the general asset type.
What this data cannot support
It cannot teach a robot to handle goods. Listing photos show static states with no joint trajectories, grasp attempts, contact forces or tactile signals. A grade of "good" does not tell a gripper how fragile a seam is. Humanoid manipulation of used goods needs robot or teleoperation demonstrations and, for fragile items, force-aware capture; recognition and attribute data alone do not provide control.
It also lacks manufacturing context: process, tolerances and original specifications are not part of a resale record. Condition labels reflect a seller's perspective, not an engineering inspection. And a resale company's mix reflects its own channel, such as fashion or electronics, so coverage is likely to be skewed. Evaluate on held-out categories and suppliers rather than random splits of the same listings.
Questions to ask suppliers
- What categories and approximate share of items does your inventory history span, and how has the mix changed?
- How are condition grades defined, who assigns them, and has agreement between graders been measured?
- Were photos taken by staff to a standard, by consumers, or both?
- Do records link back to sellers or buyers, and how are those links removed?
- What terms did consumers accept when submitting items or photos, and do they cover third-party licensing?
- Are defects localized in images or described only in text?
- Do you keep intake, refurbishment or repair images and notes?
- Is there an outcome record per item, and is it reliable?
- Can you share a stratified sample across categories and grades for testing before licensing?
Discuss your requirements
If you want condition-graded images of used goods, describe the categories, grading needs and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX reviews the requirement and follows up about potential sources. Start at the humanoid training data guide or the product and object data hub. The general category is described under product catalogs and descriptions.
Discuss your humanoid training-data requirements with SourceX
Sources
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — catalog-product images, metadata and 3D models; format reference only.