Short answer
Foreign object detection data shows unwanted material in or on a product or work area, usually as optical or X-ray images with a disposition. Buyers look to food and aerospace manufacturers with machine-vision or X-ray systems. Such data can train and evaluate detection, but it contains no robot actions, so it cannot teach a humanoid to remove or avoid the object.
What makes contamination its own target
Most defect datasets describe flaws of the product itself: a scratch, a void, a missing component. Contamination is material that should not be there at all. In a food plant it might be plastic, glass, metal fragments or bone. In aerospace work, "foreign object" usually means a tool, fastener or debris left in an assembly or on a surface. The visual appearance varies enormously because the contaminant is, by definition, arbitrary.
That arbitrariness shapes the data you need. A model that has seen only three kinds of contaminant may miss a fourth. The label that matters is often not "contaminant type X" but "this should be rejected," plus what was actually found when the item was inspected by hand.
For the general landscape of defect images, start from sourcing defect image datasets. The section overview is in the inspection, defect and failure data hub, and the full map is in the humanoid training data guide.
Where the images come from
Contamination images are generated at points where a plant tries to catch problems before shipment:
- Optical reject stations on conveyors, where a camera flags an item and a diverter removes it. See in-line machine vision archives.
- X-ray inspection, which reveals dense contaminants inside sealed products. Image conventions are covered in X-ray and CT inspection images.
- Manual checks and swab or sieve results recorded by quality staff, which may link to an image.
- Complaint investigations, where a returned item is photographed and the contaminant identified.
- Cleanliness checks in assembly areas, recorded as photos with a pass or fail.
Typical suppliers are food and beverage manufacturers, aerospace manufacturers and maintenance providers, and other plants with machine-vision or X-ray equipment. Each may hold different content, with different confidentiality risks.
A public reference, and where it does not apply
GDXray is a public collection of 19,407 X-ray images in five groups: castings, welds, baggage, natural objects and settings, with annotations as bounding boxes or labels (Mery, 2015). It does not cover food contamination or aerospace foreign object debris specifically. Use it only as a capture, labeling and format reference: it shows how X-ray images can be organized by group and annotated with boxes, and it gives a way to discuss annotation conventions with a supplier. It is not evidence of what a food or aerospace archive contains. Check the dataset's own terms before using it for anything beyond research and education.
What the data can support
- Detection and localization models that find contaminants in the image, trained on labeled positives and many negatives.
- Anomaly detection where contaminants are too varied to enumerate; see anomaly detection versus supervised defect data.
- Evaluation of false alarms using negative-heavy archives that reflect real production.
- Disposition reasoning, such as predicting reject versus release, if dispositions are linked.
Limits: what this data cannot support
It cannot train manipulation. Detecting a contaminant is a perception task. Picking out a fragment, wiping a surface or retrieving a tool from a cavity requires robot or teleoperation data, including contact, access and force information that images do not carry.
It rarely covers rare events well. Real contamination is infrequent by design, so archives are dominated by clean items. See rare defects and class imbalance for how to plan around that.
It may be tied to one line, one product and one imaging setup. X-ray images depend on source settings, product density and belt geometry. A model trained on one machine may not carry over without relabeling or adaptation.
Planted contaminants used to test systems are not the same as natural ones. Ask which is which, because a model that learns the look of a planted test piece (a standard-shaped metal or plastic target, for example) may learn the target rather than real contamination.
It does not cover hazards in the surrounding environment, such as spills or debris on a floor. That is a different target, discussed in environmental hazard images, and it needs its own imagery from the viewpoints a mobile robot would have.
Illustrative evaluation package
The following is an illustrative package outline with synthetic values, not an existing dataset.
| Component | Example content | Why it matters |
|---|---|---|
| Images | Optical and X-ray frames from line_A and line_C | Separates modalities and lines |
| Label | contaminant class, bounding box, "none" | Includes negatives |
| Disposition | rejected / released / reworked | Links image to outcome |
| Origin | natural / planted test piece | Avoids mixing test pieces with real events |
| Imaging metadata | source setting, belt speed, lighting | Supports domain-shift analysis |
| Split key | lot_ID, production date | Prevents lot leakage |
Quality and acceptance criteria
Criteria are yours to set. Useful tests on a sample:
- A reviewer can find the contaminant in a random sample of positive images.
- Negatives are labeled as reviewed, not merely "not flagged."
- Bounding boxes or masks follow a written convention.
- Planted and natural contaminants are separable by a field.
- Splits are by lot or line, not by random image; domain shift across lines explains why.
Permissions and provenance
Contamination records are sensitive. They can reveal product design, supplier problems, customer parts and incidents a company may not want associated with its name. Images may include customer part numbers or identifying marks. Ask whether any content belongs to a customer or regulator, whether recalls or investigations are involved, and who can approve licensing. See confidentiality in defect data. Redaction of marks does not by itself settle rights. This is not legal advice.
Questions to ask suppliers
- Which modalities exist (optical, X-ray, other), and from which equipment?
- Are contaminants confirmed by hand inspection, or only flagged by the machine?
- Are planted test pieces mixed in with natural events?
- What proportion of archived images are negatives, and were they reviewed?
- How are dispositions and lot identifiers linked to images?
- Have settings or equipment changed over the archive's life?
- Can images be exported in original, uncompressed form with metadata?
- Does any image reveal customer parts, designs or markings?
- Who can authorize licensing for content tied to customer or regulator interactions?
Discuss your requirements
If you need contamination or foreign object images, describe the product type, modality and evaluation needs to 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. The catalog page for manufacturing quality records describes a related data kind.
Tell SourceX what your evaluation set needs to cover
Sources
- GDXray: The Database of X-ray Images for Nondestructive Testing — X-ray image groups and annotation formats, as a reference only.