Property Inspection Photos Linked to Findings for Underwriting and Claims Models

This page covers the photo-plus-finding pair. If you mainly need the narrative report text, see inspection report data for AI training; for the wider image category, start at the image data hub.

Why photo-finding pairs matter more than raw property images

Paired data matters because the finding is the label, and an unlabeled inspection photo teaches a model almost nothing about condition. Research on roof assessment frames inspection as a slow, costly and sometimes hazardous human task that insurers want to automate from imagery [1]. That automation only works when the training signal says what the inspector concluded about the specific thing in the frame.

Three model families consume these pairs. Condition and hazard classifiers (for example, "double-tapped breaker", "missing GFCI near sink", "active roof leak staining") need image-level or region-level labels. Vision-language models fine-tuned to draft findings need the finding text as the target caption, which is covered in more depth in using inspection notes and adjuster comments as image text. Underwriting automation models need the finding rolled up to a property-level decision, such as refer, require repair or accept.

Overhead imagery is a different product. If your target is roof geometry or hail from aerial or drone captures, the roof condition and hail damage imagery guide is the better fit; ground-level photos see what aircraft cannot, such as the electrical panel, water heater venting and under-sink plumbing.

Underwriting inspections versus claims inspections

Underwriting and claims inspections produce different labels, so treat them as two datasets even when one vendor holds both. Mixing them silently corrupts the target variable.

• Underwriting inspections (new business or renewal) look for hazards and maintenance: roof age and covering, wiring type (knob-and-tube, aluminum branch circuits), panel brand, trampoline or pool fencing, wood stoves, handrails. Findings are usually hazard codes with a required-repair flag and due date.
• Claims inspections document a loss: water intrusion source, mold extent, fire and smoke zones, wind or hail impact. Findings tie to a cause-of-loss code, affected rooms and an estimate. For estimate line items and adjuster field notes, see property claim estimates and adjuster field reports.
• Pre-purchase home inspections follow a standards-of-practice structure by system (roof, exterior, structure, electrical, plumbing, HVAC, interior) and tend to have richer narrative but less consistent severity scales.

A claims-heavy dataset over-represents damage; an underwriting dataset over-represents clean homes with a handful of hazards. Know which base rate you are buying.

What a usable photo-finding record looks like

A usable record links one finding to one or more photos by ID, carries structured codes alongside the free text, and keeps capture context without exposing the household. Most source systems store photos as attachments on a report section, so finding-level linkage often has to be reconstructed by the supplier, which is the first thing to verify in a sample.

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

{
  "inspection_id": "INSP-7f3a",
  "inspection_type": "underwriting_renewal",
  "property": {"region": "US-South", "year_built_band": "1970-1979", "dwelling_type": "single_family"},
  "finding": {
    "finding_id": "F-012",
    "system": "electrical",
    "component": "main_panel",
    "code": "ELEC-PANEL-DOUBLE-TAP",
    "severity": "repair_required",
    "action": "licensed_electrician",
    "text": "Two conductors under single breaker terminal, lower left."
  },
  "photos": [
    {"photo_id": "P-0441", "role": "context", "room": "garage", "redacted": false},
    {"photo_id": "P-0442", "role": "detail", "room": "garage", "redacted": true,
     "regions": [{"bbox": [812, 604, 140, 96], "label": "double_tap"}]}
  ],
  "capture": {"device_class": "phone", "exif_gps_removed": true, "capture_month": "2025-04"},
  "inspector": {"inspector_hash": "a91c", "certification_level": "senior"}
}

For annotation export, COCO JSON is the common interchange: its info, licenses, categories, images and annotations sections map cleanly to photo IDs, finding codes and bounding boxes, and tools such as CVAT import and export it [6]. For large deliveries, WebDataset tar shards (often around 1 GB each) keep each image beside its JSON sidecar and stream well at training time [7].

In-home privacy: what interior photos expose and how to redact it

Interior photos are a privacy problem distinct from report text, because a camera inside a home captures far more than the finding. Define redaction rules per object class before any sample is pulled, and record which method was applied.

What typically appears in frame:

• People: occupants, children, the inspector's reflection in mirrors and appliance doors.
• Documents: mail, bills, prescription bottles, calendars and whiteboards with names or schedules.
• Identity and location cues: house numbers, license plates in driveways, street signs, mailbox names, EXIF GPS coordinates. See what EXIF metadata to strip and keep.
• Sensitive possessions and status: family photos, religious items, firearms, medical equipment, safes and valuables.

Faces are the highest-risk class. Illinois BIPA regulates collection, retention and disclosure of biometric identifiers [3], so any pipeline that runs face geometry on interior photos, including for blurring, needs a legal review of whether it creates biometric data. Under California's CCPA, data only counts as deidentified if the holder meets specific conditions, including reasonable measures against re-identification [4]. Memorization is also real: diffusion models have been shown to regenerate individual training images, including photos of people [5], which argues for redaction before training rather than relying on output filters.

Practical rules buyers can ask for: blur or mask faces and full bodies, documents and screens, house numbers and plates; strip EXIF GPS and device serials; coarsen capture date to month; replace addresses with region bands; and drop images where redaction would remove the finding itself.

Label quality, coverage and bias in inspection findings

Inspection labels are noisy because inspectors differ in thresholds, templates change, and findings reflect what the inspector chose to photograph. Audit before you train.

Underwriting outcomes such as non-renewal can encode past decisions rather than true risk; the guide to historical decision bias in operational labels covers that audit. Ask suppliers for a written data statement describing who inspected, under which template, in which period, so reviewers can judge generalization [8]. Rare hazards (active arcing, structural cracking) will be thin; the rare defect coverage guide covers pooling and synthetic augmentation.

Rights, consent and insurance-regulator expectations

Rights in inspection photos are layered: the homeowner owns the property, the inspector or firm usually owns the photos, and the client (buyer, agent or insurer) may hold contractual use rights. Confirm the chain before evaluating samples.

Questions for counsel and the supplier:

• Do the inspection agreement and privacy notice permit use of photos and findings for model development, or only for the report?
• Did the homeowner or occupant consent to interior photography, and on what terms?
• Are any photos of third-party artwork, logos or identifiable people that need releases? See model and property releases for AI training images.
• If an insurer will deploy the model, can the data's provenance satisfy the NAIC Model Bulletin, a guidance document (not a model law) that a growing number of states have adopted as of October 2026 and that expects documented governance over AI systems and third-party data [2]? The NAIC bulletin guide for insurers goes deeper.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Sample request template for inspection photo-finding data

A precise request gets better matches than a category name. Describe the data, not the companies you think hold it.

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

How SourceX approaches inspection photo sourcing

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing and ongoing purchases. Nothing is held in stock, and a request does not guarantee a match. Buyers describe the data; SourceX looks for US businesses that hold it, and every release is approved by the supplying company.

Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, allowed uses, term and delivery. Names, emails, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. You can describe your inspection photo requirements to SourceX using the template above; related categories are listed under images and inspection photos, underwriting files and the insurance buyers page.

Request property inspection photos with findings

SourceX finds US companies that hold the inspection photos and findings you describe, assesses data and licensing permissions, and agrees pricing and allowed uses in a license before anything is delivered. Nothing is contracted until a supplier agrees. Start a buyer request at sourcex.si/buyers.