Glass and Transparent Part Defect Data: Sourcing Images of Clear Materials

The buyer's problem

Place a clear bottle on a bench and photograph it with a room light: you mostly see the background through it, plus reflections. A hairline crack, a seed bubble or a thin stress line may be invisible. Factories solve this with deliberate illumination, and the images they keep reflect those choices. The same defect can vanish under bright-field backlight and stand out under dark-field side light.

So a transparent-part dataset is a dataset of image-plus-illumination pairs. If the archive stores only one lighting mode, it supports only the defects that mode can show. If it stores several, labels have to say which mode each defect was marked in.

Defect classes in clear materials

• Inclusions and bubbles. Seeds, blisters, stones, unmelted particles, black specks in clear plastic.
• Surface damage. Scratches, scuffs, chips, checks and cracks, edge damage on cut glass.
• Optical distortion. Cords, striae, waves, warp, lens-like distortion, haze or cloudiness.
• Molding artifacts in plastics. Flow lines, sink marks, weld lines, short shots.
• Contamination. Dirt, residue, fingerprints, coating faults.
• Dimensional faults. Wall thickness variation, neck or finish defects on containers.

Where the target is geometry rather than appearance, see 3D surface inspection data. Many of these are optical rather than geometric, so the label may describe an effect on light rather than a visible shape. See pixel-level defect segmentation for what mask labels can and cannot capture when a defect has soft or distributed boundaries.

Illumination decides what the image can show

The practical rule is that the metadata is as important as the pixels. Imaging conditions metadata lists what to demand; for transparent parts, add illumination mode, backlight geometry, polarizer settings and exposure per image.

What public transparent-object work shows

ClearGrasp addresses depth estimation for transparent objects so a robot can grasp them, using a synthetic training set of more than 50,000 RGB-D images and a real-world benchmark of 286 RGB-D images (Sajjan et al., 2019). It shows that clear objects are hard for depth sensors, not how to find defects on them. It is a perception-and-grasping reference that does not cover glass or clear-part inspection, so it is relevant here only for the difficulty of the imaging problem.

MVTec AD 2 includes transparent and reflective objects among its harder scenarios, and its test_private_mixed split was captured under varied lighting conditions (Heckler-Kram et al., 2025). It is a useful reference for designing an evaluation set that varies lighting on purpose. It is a reference that does not cover glass or clear-plastic production inspection, and its CC BY-NC-SA 4.0 license means commercial use needs a license check; this is not legal advice.

Where the data lives

Container-glass makers, flat-glass processors, optical and lighting-lens molders, medical-device plastics producers and labware manufacturers all inspect clear parts. Expect machine-vision archives from fixed stations with known lighting, usually retained as reject images with a few good ones. Hand inspection under inspection lights leaves notes, not images. Customer returns and complaints sometimes come with photos, which are taken under uncontrolled light and are best treated as a separate stratum; see warranty and field failure data.

Foreign material in clear products or packaging is a related but distinct task discussed in foreign object detection.

Illustrative image record for a transparent-part package

# Illustrative only - synthetic ids and values
package: transparent_part_defects_v0
part_family: clear_container_type_M
image_record:
  image_id: img_004417
  part_id: part_88231
  illumination: dark_field_left    # backlight | dark_field_left | polarized | coaxial
  exposure_note: recorded
  camera_id: cam_C
  lens_note: telecentric        # recorded as supplied
labels:
  - defect: chip
    box: [x, y, w, h]
    visible_in_modes: [dark_field_left]
  - defect: seed_bubble
    box: [x, y, w, h]
    visible_in_modes: [backlight, dark_field_left]
disposition: reject
split_key: part_id then production_lot
excluded: robot actions, joint states, forces, depth-grasp data

Field names are examples. Per-mode visibility is the unusual and valuable column, since it lets you test whether a defect is findable by one mode alone.

Quality and acceptance pointers

• Mode coverage. Count images per illumination mode and per defect class.
• Reference parts. Ask for good parts imaged in every mode, not only defective ones; see golden samples and reference images.
• Cleanliness. Fingerprints and dust on clear parts produce false defects; ask how often parts were cleaned before imaging.
• Mask versus box. Distributed defects such as haze may not fit boxes.
• Hold-outs. Split by part and lot, and by lighting setup if your deployment will change it.

What this data cannot support

Defect images of clear parts contain no robot grasps, forces or trajectories. Transparent-object manipulation, such as picking a glass tube or handling labware, is a different problem: depth sensing fails on clear surfaces, and training requires robot or teleoperation data with synchronized sensors, not inspection archives. The object recognition vs manipulation data page explains the gap.

Inspection images also cannot show whether a flaw affects strength or sealing unless outcomes such as breakage tests are linked. Clear-part archives are dominated by good parts, and light-sensitive defects may be underrepresented if only one illumination mode was used. A model trained on one setup's optics can fail on another.

Questions to ask suppliers

1. Which illumination modes were used, and is the mode stored per image?
2. Did lighting, camera or optics change over the archive?
3. Are good parts imaged in the same modes as rejects?
4. Were defects marked in one mode or across several?
5. How were doubtful, faint defects labeled, and by whom?
6. Were parts cleaned before imaging, and are contamination artifacts flagged?
7. Which defect classes have real examples and which only a handful?
8. Are breakage, leak or sealing test results linked to any images?
9. What customer designs, logos or markings appear, and who may license them?

Permissions and provenance

Clear containers and optical parts often carry embossed customer logos or follow customer-owned designs. Supply agreements may limit the use of records. Ask who owns the geometry, whether any mold or design is proprietary to a customer, and whether restrictions follow from regulated products such as medical or food-contact plastics. This is not legal advice; removing a logo from an image does not by itself resolve a rights question.

Discuss your requirements

If you need glass or transparent-part defect imagery, describe the material, defect classes, illumination modes 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 images and inspection photos.

Explore sourcing and licensing options with SourceX

Sources

• ClearGrasp: 3D Shape Estimation of Transparent Objects for Manipulation — shows how hard transparent objects are for depth sensing; not a defect dataset.
• The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection — transparent and reflective scenarios and a varied-lighting test split as an evaluation-design reference.