Field Photos Matched to Parts Catalogs for Visual Part Identification

Why field-to-catalog matching needs its own data

Field-to-catalog matching is a cross-domain retrieval problem, so clean product shots alone will not train it. A technician's phone photo of a seized bearing, a corroded hydraulic fitting or a belt with the printed part number worn off looks nothing like the white-background render on a distributor's product page. General vision-language models such as LLaVA are built from broad image-text pairs [1], and those pairs rarely include a 6205-2RS bearing photographed inside a gearbox under a work light.

This differs from adjacent problems. Part cross-reference matching resolves part numbers and descriptions in text. SKU recognition on retail shelves handles front-facing packaged goods, where benchmarks such as SKU-110K stress dense, look-alike products on shelves [4]. Industrial parts are usually unpackaged, often installed, and distinguished by dimensions, thread pitch, material or revision letter that may not be visible.

Ground truth: what counts as a confirmed match

The strongest label is the SKU a transaction or maintenance record proves was supplied or installed for the photographed part. Annotator guesses from images are weak labels for this task, because two parts that look identical can differ in bore size or pressure rating. Ask suppliers how each match was confirmed and keep that as a field.

A useful confidence ladder, from strongest to weakest:

• Order-confirmed: photo attached to a quote request or return, linked to the invoiced order line for the replacement part.
• Work-order-confirmed: photo on a maintenance work order whose parts-consumed line records the installed SKU, as in CMMS records (maintenance work order datasets).
• Nameplate-confirmed: a legible manufacturer part number or nameplate in the photo, matched to the catalog.
• Expert-assigned: a counter specialist or engineer assigned the SKU without a transaction record.

Watch for substitution. A distributor may ship an equivalent from another manufacturer when the original is unavailable, so the invoiced SKU can differ from the part in the photo. Records should carry both the requested part and the supplied part, with a flag for approved substitutes.

Hard negatives and look-alike parts

Hard negatives (catalog items that look like the correct answer but are wrong) teach a part identification model more than random negatives do. Contrastive embedding training benefits from mined hard negatives, but mining can surface false negatives, items that are actually valid matches, which must be filtered out [2]. In industrial catalogs this is a real problem: interchangeable parts from different brands, superseded part numbers and kit-versus-component listings can all be legitimate answers.

Specify hard negatives explicitly:

• Same family, different size (a 6205 versus 6206 bearing, a -08 versus -10 JIC fitting).
• Same dimensions, different manufacturer, where interchangeability is not approved.
• Same part number, different revision or seal type (2RS versus ZZ).
• Superseded numbers, recorded as equivalence rather than as negatives.

An interchange or supersession table delivered with the data lets you separate true negatives from acceptable alternates. Without it, retrieval metrics will punish correct answers.

Field photo conditions to require

Capture conditions should mirror deployment, because the hardest failure mode is a model that works on clean reference images and fails on the photos users actually send. Request a documented distribution across:

• Wear and contamination: grease, rust, paint overspray, heat discoloration, broken flanges.
• Viewpoint: installed in place, partial views, oblique angles, multiple photos of one part.
• Lighting and device: phone flash in enclosures, low light, motion blur, varied phone models.
• Scale cues: whether a caliper, ruler or hand is in frame, and whether measured dimensions are recorded.
• Text visibility: nameplate legible, partly legible or absent.

Curation matters as much as volume; studies of image-text datasets show that filtering choices change downstream model quality substantially [3]. Ask how blurry, duplicate and wrongly attached photos were removed, and request the rejection counts, not just the final set. For deeper guidance on labeled inspection imagery, see industrial defect image datasets and images and inspection photos.

An illustrative record schema

A record should link every field photo to one confirmed SKU, its catalog assets, the evidence for the match and its rights status. The example below is a starting point for a request.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "record_id": "fp-000184",
  "field_images": [
    {"file": "fp-000184-1.jpg", "view": "installed_oblique", "exif_stripped": true},
    {"file": "fp-000184-2.jpg", "view": "removed_nameplate", "exif_stripped": true}
  ],
  "capture_context": {"asset_class": "conveyor_gearbox", "condition": ["grease", "seal_damage"]},
  "requested_sku": "MFR-A 6205-2RS",
  "supplied_sku": "MFR-B 6205-2RSH",
  "substitution": {"is_substitute": true, "approved_interchange": true},
  "match_evidence": "work_order_parts_line",
  "catalog_assets": {
    "images": ["cat-6205-2RSH-front.png"],
    "image_rights": "distributor_owned_photography",
    "spec_text": "Deep groove ball bearing, 25 mm bore, 52 mm OD, 15 mm width, double rubber seal"
  },
  "attributes": {"bore_mm": 25, "od_mm": 52, "width_mm": 15, "seal": "2RS"},
  "hard_negative_skus": ["MFR-A 6205-ZZ", "MFR-A 6206-2RS"],
  "split": "test",
  "split_group": "site_017"
}

Note split_group: photos from one site, one technician or one asset tend to look alike, so split by group, not by record, or your test set will overstate accuracy.

Rights on catalog images and spec text

Catalog images often carry different rights from field photos, so check training permission for each asset type separately. Manufacturer product photography and datasheets may be licensed to a distributor for display on its storefront, which does not by itself grant the distributor a right to license them for model training. Distributor-shot photography, distributor-written descriptions and customer-submitted field photos each need their own ownership and consent check.

Dataset metadata is not a reliable substitute: an audit of popular dataset hosting sites found licenses omitted for over 70% of datasets and miscategorized in over 50% [6]. Ask for a per-asset rights field, as in the schema above. The guide to licensing multimodal records from several rightsholders and the image licensing terms guide cover how to structure that. For catalog-only data, see product catalogs and descriptions.

Privacy in field photos

Field photos contain more personal data than buyers expect, including technicians' faces and hands, badges, customer site names, whiteboards and GPS coordinates in EXIF metadata. Research on large image datasets shows that faces and identifying documents survive automated filtering [5]. Require EXIF and GPS stripping, face and badge blurring, and redaction of customer names in work order text before delivery.

SourceX removes or replaces personal details such as names, emails, phones and account numbers before delivery, records the method and checks a sample; no method is perfect. For images with people in frame, see face data consent and anonymization.

Evaluating a sample before you license

A sample review should test whether matches hold up, not just whether photos look realistic. A short acceptance checklist:

• Pull 50 records at random and verify the SKU against the stated evidence type.
• Check the substitution rate and whether interchange tables explain it.
• Confirm hard negatives exist for each family, and that none is a valid alternate.
• Measure the share of photos with a legible part number; if that share is much higher than in your real traffic, the model may lean on reading text rather than appearance.
• Confirm group-based splits and no duplicate photos across splits.
• Confirm per-asset image rights and EXIF removal.

If you plan a held-out benchmark, keep it separate from training supply; see private multimodal evaluation sets. The multimodal data hub and the multimodal specification template help turn this into a written request.

Where this data comes from

This data sits in operational systems: distributor quote and returns workflows, e-commerce "identify my part" uploads, CMMS work orders and field service apps. SourceX sources operational datasets from US companies on request, including engineering records and documents, and manages the commercial process through licensing. Nothing is held in stock and a request does not guarantee a match. Buyers in industrial distribution can describe the photo-to-SKU data they need without naming suppliers; SourceX looks for US businesses that hold it.

Request industrial part identification data

Describe the parts families, photo conditions, evidence types and rights you need, and SourceX will look for US businesses that hold that data. Every dataset is rights-reviewed and delivered under a license defining records, uses, term and delivery, with each release approved by the supplying company. Start a buyer request.