Short answer
Building and infrastructure defect images are typically held by inspection firms, engineers and property or asset operators who photograph concrete, steel, roofs and facades in the field. Conditions are uncontrolled, defects often co-occur, and scale, location and element identifiers decide whether an image is usable. Report text and severity ratings add value. The images do not train robot movement or manipulation.
How field imagery differs from factory imagery
A factory camera sits at a known distance, under known light, looking at a part presented in a known pose. A bridge inspector photographs from a gantry, a drone, a ladder or the ground, at whatever distance access allows, in whatever weather and sun angle the day offers. The same crack is a hairline in one frame and a thick dark band in another, and shadows, stains, efflorescence, graffiti and vegetation produce lookalikes.
For a buyer, that means more of the burden shifts onto metadata and labels. Without a scale reference, you cannot say how wide a crack is. Without an element identifier, you cannot tell which pier or slab you are looking at. Without the inspector's rating, you do not know whether the defect mattered.
Defect types and why they co-occur
Concrete and masonry defects commonly include cracks, spalling, exposed reinforcement, corrosion staining and efflorescence. Steel structures add corrosion, section loss, coating failure and deformation; roofs and facades add ponding, membrane damage, water ingress and sealant failure.
The CODEBRIM dataset of bridge images was built for multi-target classification of five concrete defect types that often co-occur on the same patch of surface, and the authors' best multi-target test accuracy was up to 72% (Mundt et al., 2019). Two lessons transfer to purchasing. Labels should be multi-label, since a single-class tag misrepresents images where spalling exposes corroded bars. And a headline score on a public set should not be read as what a model will achieve on your assets, whose finishes, materials and conditions will differ.
Who holds the images
| Supplier | What they hold | Common limit |
|---|---|---|
| Engineering and inspection firms | Photo logs and reports for client assets | Client owns the report and often the photos |
| Municipalities, transit and utilities | Periodic inspection photos and condition ratings | Public-records and security considerations |
| Property operators and facility managers | Condition assessments, work orders with photos | Photos taken ad hoc by maintenance staff |
| Drone service providers | Survey imagery with flight metadata | Coverage tuned for modeling, not defect labeling |
| Insurers and adjusters | Claim photos | Privacy and claim confidentiality |
Third-party firms often generate much of the professional imagery, and the rights picture is described in third-party inspection firms as suppliers. For catalog context on related records, see construction project records.
Metadata that makes field images usable
- Scale. A ruler, crack comparator or known object in frame, or a calibrated camera distance.
- Element and location. Structure, member, face and position, preferably from the inspection's own scheme.
- Date and condition cycle. Allows change over time to be studied across successive inspections.
- Capture method. Handheld, pole, drone, rope access, with camera and approximate distance.
- Weather and light. Wet surfaces darken concrete and hide or reveal cracks.
- Inspector rating. The condition state or priority the engineer assigned, with the manual it came from. This page cites no inspection manual or code, so take rating scales and definitions from the supplier's own documentation.
- Report link. The narrative that explains the photo; see inspection report text.
Photos in a typical report are selected to illustrate findings, so they under-represent routine good surfaces. Ask whether the full photo log or only report figures are retained.
Illustrative package structure
# Illustrative only - synthetic ids and values; no real asset is described
package: structure_defect_images_v0
asset_type: concrete_bridge_deck # example category
records:
- image: img_30192.jpg
asset_id: asset_017 # pseudonymous
element: deck_span_3_underside
capture: {method: handheld, distance_note: "approximate", scale_ref: ruler}
date: 2025-xx-xx
weather_note: dry
labels: [crack, efflorescence] # multi-label
mask_file: img_30192_crack.png # optional
inspector_rating: "condition state per owner's manual"
report_ref: rpt_2025_041
split_key: asset_id then inspection_cycle
excluded: robot actions, joint states, forces
All values are examples. The buyer decides whether masks or labels suffice and whether severity should be a requirement.
Quality and acceptance pointers
- Class balance by asset. Check that defect types are not concentrated on a single asset.
- Scale availability. Ask what share of crack images carry a usable scale reference.
- Label noise. Staining and construction joints are routinely mistaken for cracks; sample them.
- Split by asset. Different photos of the same pier are near-duplicates; hold out whole assets, and when the claim requires it, whole regions or climates.
- Rating consistency. Ratings from different inspectors and years may follow different manuals.
None of these has a universal pass mark. Match thresholds to your use. For the broader issue of rating outcomes, see defect severity and disposition.
What this data cannot support
Field defect photos describe what a surface looked like at the time of inspection. They contain no robot joint states, actions or forces, so they cannot train a humanoid or any robot to climb, reach, tap-test or repair; that requires robot or teleoperation data, and active inspection by humanoids explains how passive images differ from interactive inspection. They do not carry structural capacity either, so a crack image alone does not say whether an element is safe.
Coverage is skewed. Inspection photos favor visible, accessible surfaces and concerning findings, while hidden members and well-behaved elements are under-represented. Climate, materials and construction practices vary by region, and models may transfer poorly across them. Serious structural failures are rare and are seldom in any archive. Related damage forms are discussed in equipment wear and damage images and environmental hazard images.
Questions to ask suppliers
- Who owns the photos and the underlying reports, and is the client's consent needed to license them?
- Are full photo logs retained, or only the images embedded in reports?
- What scale references, element identifiers and dates are stored with each image?
- Which inspection manual or rating scheme defined condition states?
- Were defects labeled by engineers, technicians or after-the-fact annotators?
- Were photos taken by drone, handheld or rope access, and does that vary by asset?
- Do the images include people, vehicles, license plates, signage or addresses?
- Are any assets critical infrastructure with security restrictions on imagery?
- Are repair records available for assets later fixed?
Permissions and provenance
Field images raise their own questions: property owner consent, public-agency records rules, critical-infrastructure sensitivity, bystanders and vehicles in frame, and drone flight authorization. Ask the supplier who can grant a license and what the inspection contract says. This is not legal advice. Cropping or blurring faces and plates does not by itself resolve a permission question or a security concern, and asset identity may still be inferable from the structure.
Discuss your requirements
If you need building or infrastructure defect images, describe the asset types, defect classes, label level and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match. Continue with the inspection data hub, the humanoid training data guide, or the catalog entry for inspection reports.
Explore sourcing and licensing options with SourceX
Sources
- Meta-learning Convolutional Neural Architectures for Multi-target Concrete Defect Classification with the COncrete DEfect BRidge IMage Dataset — bridge-image dataset with co-occurring concrete defect types and reported multi-target accuracy.