Short answer
Paint and coating defect data is typically held by finishing lines, auto body operations and coating applicators that photograph or scan finished surfaces. Gloss, metallic flake and color change how each defect looks, so lighting geometry, finish type and repair outcomes must travel with the images. The data can support surface inspection models, not painting robots or humanoid control.
Why finishes are hard to image
A scratch on matte gray plastic and a scratch under a high-gloss clear coat are different visual events. On a mirror-like surface the camera mostly sees reflections, and a dent shows up as a distortion of the reflected pattern rather than as a patch of different color. Metallic and pearl finishes shift hue with the viewing angle, so the same flake cluster looks like a defect from one direction and normal from another.
For a buyer this means a painted-surface image is not self-describing. The lighting arrangement, camera angle and illumination type are part of the evidence, and an archive that dropped them has lost something a model needs. The general metadata list is on imaging conditions metadata; this page covers what finishing adds.
Defect classes a finishing archive may hold
Classes differ by process stage and by supplier, so ask for the supplier's own list rather than assuming a standard one.
- Application defects. Runs and sags, orange peel, dry spray, thin or missing coverage, craters, solvent pops.
- Contamination. Dirt or fiber inclusions, fisheyes, overspray, sanding marks showing through.
- Color and gloss deviation. Mismatch between panels, mottling in metallic finishes, gloss variance.
- Damage after coating. Scratches, chips, dents and handling marks; see vehicle body damage images for field damage rather than line defects.
- Edge and masking faults. Bleed under masking, missed areas, uneven film at edges.
Bare-metal surfaces before coating are covered in metal surface defects. Classes tend to suit different imaging, for example low-angle light for texture, diffuse domes for color and structured or patterned light for gloss form; confirm the setup with the supplier. A single image per panel rarely serves all of them.
What published datasets can and cannot tell you
The MVTec Anomaly Detection Dataset contains 5,354 high-resolution color images in 15 categories, five of them textures, with pixel-precise ground truth for all anomalies and 70+ defect types; its training set is defect-free only, and it is released under CC BY-NC-SA 4.0 (Bergmann et al., 2021).
It is a reference that does not cover painted or coated automotive or industrial finishes, so treat it only as a capture, labeling and format reference: it shows what pixel-level ground truth and a defect-free training protocol look like. The non-commercial license also marks a typical gap between academic sets and what a company can lawfully use in a product; this is not legal advice and your counsel should read the actual terms. Nothing in it tells you how your clear coat or metallic base behaves.
Where finishing data lives
| Supplier type | What they may retain | Common gap |
|---|---|---|
| Automotive body and paint shops | Station images, inspection-tunnel captures, defect marks | Marks recorded on a tablet or sticker, not tied to image pixels |
| Industrial and appliance finishing lines | Spot-check photos, quality-hold records | Photos taken by hand under varying light |
| Powder and wet coating applicators | Thickness readings, rework logs | Few images of the defects themselves |
| Refinish and repair operations | Before-and-after pictures | No consistent camera position |
Automated inspection tunnels, where present, are often the most usable source because lighting and geometry are fixed. Hand-photographed spot checks are cheaper to find and harder to use.
Link images to what happened next
An image of a defect becomes more valuable when it is tied to the decision. Did the panel get spot-repaired, polished, resprayed or scrapped? A craters-and-dirt image that triggered a polish is a different training signal from one that triggered a full respray. The decision structure is covered in defect severity and disposition, and the reference condition for a finish in golden samples and reference images.
Color deserves a note. Color-matched finishes are specified as formulas and tolerances, and a defect label such as "off-color" is meaningful only relative to a master panel. Ask which master panel and which tolerance applied on which date.
Illustrative acceptance table for a finish-defect package
| Check | Illustrative threshold style | Why |
|---|---|---|
| Lighting record present per image | Required field, no blanks | Appearance depends on geometry |
| Finish type recorded | solid / metallic / pearl / clear-over-base | Defect appearance differs by finish |
| Panel and part IDs | Present for all images | Needed for splits and outcome joins |
| Defect marks tied to pixels | Box or mask coordinates, not a position description | Enables localization labels |
| Repair outcome | Present for a stated share of defect images | Enables severity learning |
| Split key | Panel or vehicle, then line and week | Avoids near-duplicate leakage |
The thresholds are examples the buyer sets, not universal pass marks. Values such as share of defect images with repair outcomes depend on the model and the task.
Limits of finishing-line data
This data cannot support manipulation or control. It contains no painting trajectories, spray gun poses, joint states or forces, so it is not training data for a robot that sprays, sands or handles finished panels; that needs robot or teleoperation data. It also does not describe the coating process itself unless you also obtain process records, such as booth conditions, which are a separate asset.
Domain shift is a real concern: a model fit on one booth's lighting and one color family can degrade on another, which domain shift across lines discusses. Rare defect classes will be sparse, and the most severe ones are reworked before they reach a camera at the final station. Hand-painted prototype or customized finishes are unlikely to appear at all.
Questions to ask suppliers
- Are the images from fixed inspection stations or hand-held spot checks?
- What lighting types and angles were used, and did they change over time?
- Which finish types and color families appear, and how many of each?
- Were defects marked on the image, on the panel or only in a log?
- Are repair and scrap outcomes available per defect?
- Which master panel or color standard defined "off-color", and when was it updated?
- Were images compressed or cropped before storage?
- Which customers' parts, colors or logos appear, and who may license them?
- Are false rejects and later-overturned calls identifiable?
Permissions and provenance
Finishing lines frequently paint for customers who own the product design and sometimes the color formulation. Images of unreleased vehicles or products can carry confidentiality terms. Ask whether supply agreements restrict use of inspection images, and whether license plates, VINs or facility details appear. This is not legal advice; blurring such details does not by itself resolve a permission question.
Discuss your requirements
If you need finish-defect imagery, describe the finish types, defect classes, lighting setup 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. You can also browse the inspection data hub and the humanoid training data guide, or read about images and inspection photos.
Explore sourcing and licensing options with SourceX
Sources
- The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection — public anomaly dataset used as a capture, labeling and licensing-scope reference only.