Short answer
Returns processors and retailers photograph and grade large volumes of returned items, which makes their archives a source of condition labels such as new, open-box, used or damaged. These support condition classification and failure analysis. Grade definitions differ by seller, photos can carry customer data, and grades record a business decision rather than a physical measurement, so each label scheme must be mapped and audited.
What a condition grade is
A condition grade is an inspection label from reverse logistics. A worker looks at a returned item, checks it against a checklist, photographs it, and assigns a grade that decides where the item goes next: restock, refurbish, discount, liquidate or scrap. The grade therefore encodes two things at once: what the item looks like and what the business chose to do about it.
That dual nature is what makes these archives useful and tricky. The label is rich with operational meaning, yet the same worn corner might be "Grade B" at one seller and "Good" at another. Before you buy, you need to know the scheme, the checklist behind it, and who applied it. This page covers sourcing and acceptance; for the general ideas behind linking pictures to decisions, see inspection images linked to outcomes.
What the data can support
- Condition classification across product categories, using seller grades as labels.
- Detecting specific returns defects: missing parts, scuffs, stains, open or resealed packaging, wear marks.
- Comparing a returned item with a reference image to decide "same item, changed state".
- Failure and damage analysis, when the return reason is recorded.
- Evaluation of whether a perception model can reproduce human graders, which is a useful held-out test.
Product reference data helps the second and third uses. Catalog corpora illustrate the structure: Amazon Berkeley Objects pairs catalog images and metadata for 147,702 product listings, with 398,212 unique catalog images and up to 18 metadata attributes per product (Collins et al., 2021). That study did not involve returned items; the point is that a returns archive becomes more valuable when each item can be matched to a catalog record. See product catalog data for robot perception.
Where the archives are
| Supplier type | What they hold | Typical weakness |
|---|---|---|
| Retailers with in-house returns centers | Item photos at receiving, grade, disposition | Grade scheme tied to the retailer's resale channels |
| Third-party returns and liquidation processors | Multi-client photo archives with grading checklists | Client ownership of images and product data |
| Manufacturers with refurbishment lines | Photos tied to defect codes and repair actions | Narrow product range |
| Resale and recommerce platforms | Seller photos with condition tags | Seller-written descriptions; inconsistent photography |
| Repair and trade-in services | Intake photos and test results | Electronics-heavy; customer device data risk |
Processors often hold data on behalf of clients, so the party in front of you may not be able to license it. Ask early who owns the images and the product records; the same question arises for inspection firms in third-party inspection firms as data suppliers.
Illustrative package: graded returns photos
# Illustrative only - fields and values are examples, not an available dataset
package: returns_condition_photos_v0
source: one returns processor, 2 facilities
images: 6 fixed views per item (front, back, sides, top, label) + detail shots
grading:
scheme_version: "grading_v3, definitions attached"
grade: A | B | C | D | scrap
defects: [scuff, stain, missing_part, dent, seal_broken, odor_note]
return_reason_code: from supplier list (free text removed)
disposition: restock | refurbish | liquidate | recycle
grader_id: grader_A12 (pseudonymous)
reference: catalog_image_id and SKU where matched
audit: second-grader labels on a random 5% sample, disagreement logged
removals: customer names, addresses, order numbers, packing slips, home interiors
rights: AI training and internal evaluation; no redistribution; deletion on request
Privacy and rights considerations
Returned goods arrive with their owners' traces: shipping labels, order numbers, handwriting, photos of rooms behind a product, and electronics with personal content on screen. Photos of the returned item taken by a customer can be a different category from photos taken by the processor. Ask which are included, and under what terms each was collected.
Manufacturer defect codes, return reasons and grades may be confidential business information of a retail client. Product photos can include trademarks and packaging artwork. See product image rights for the object-side view. This is not legal advice; have counsel review the grant, and do not treat blurring or cropping as a substitute for permission.
Limitations: what returned item grades cannot do
Grades are business labels, not physical measurements. A "Grade C" tells you the item was resold at a discount, not the depth of a scratch or the cause. Seller schemes also drift over time, so labels across months may not mean the same thing.
Archives built for resale decisions do not show manufacturing context: no process data, no lot-level defect history, no in-line imagery. They also contain no robot action data; photos of items at a grading table do not teach a robot how to handle, flip or repackage them. If your goal is a robot that grades items, the images support the perception component only, and grasping and handling need robot-side data. Related handling damage is covered in handling-induced damage.
Acceptance checks
- Map each grade to a written definition and check whether examples match it.
- Have your own inspectors regrade a random sample; compute agreement with the supplier grade. Inter-inspector agreement explains how to read the results.
- Confirm views are consistent: the same fixed shot list, lighting and background.
- Check class balance by product category and grade, and that rare defect types have enough examples for your use.
- Hold out evaluation items by SKU and by processing site.
- Inspect redaction on a sample for missed labels, handwriting or customer content.
- For used and refurbished categories, see secondhand and recommerce item data for category-specific points.
Mapping grades across sellers
If you buy from more than one supplier, build a crosswalk before training. List each scheme's grades and written definitions side by side, then have an inspector regrade a shared sample of items under your own scheme. Keep the original grade as a feature and your regrade as the label. Pair this with the defect severity and disposition guidance, since a grade is partly a disposition decision. Without a crosswalk, a model will learn which seller's photos it is looking at rather than the condition of the item.
Questions to ask suppliers
- Who defined the grading scheme, and what version was in use for each image?
- Are the written checklists and photo examples for each grade available?
- How many graders worked on the archive, and was grading audited?
- Do photos include customer-taken images, or only processor-taken photographs?
- Which fields could identify a customer: labels, order numbers, packing slips, backgrounds?
- Who owns the images, product records and return reason codes, and can you grant AI-training rights?
- Are catalog images or SKUs linked to each item?
- What is the distribution of grades, and what share of items are categories you care about?
- Were any images retouched, color-corrected or resized after capture?
- Can the supplier delete records on request, and what happens to derived labels?
Discuss your requirements
If you can describe the product categories, grade definitions and permitted uses, explore sourcing and licensing options with SourceX. SourceX looks for US businesses that 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 follows up about potential sources.
Start from the humanoid training data guide or the inspection, defect and failure data hub. The catalog pages on images and inspection photos and expert annotations and labels are related.
Explore sourcing and licensing options with SourceX
Sources
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — catalog images and metadata as reference data for products.