3D Surface Inspection Data: Point Clouds and Profilometry

Why 3D data needs its own sourcing approach

A flat photograph hides some defects and exaggerates others. A shallow dent on a glossy panel may disappear under diffuse light and show clearly in a height map. A burr or a warped flange has a geometric signature that color does not carry. That is why plants use laser triangulation, structured light, stereo cameras or profilometers on lines where height and shape matter.

The data these systems produce is not uniform. One vendor stores raw profile lines, another stores fused point clouds, another stores depth images with calibration. Labels may be per part, per region or per point. A buyer who asks for "3D inspection data" without specifying these things will receive material that cannot be compared across suppliers.

For image-based counterparts, see defect image datasets and multi-view inspection imagery. The section is introduced in the inspection, defect and failure data hub, with the overall map in the humanoid training data guide.

Formats you may encounter

• Laser line profiles. A sequence of height profiles along a moving part. Needs encoder or timing data to assemble into a surface.
• Point clouds. Sets of 3D points, sometimes with color or intensity, in a sensor or part coordinate frame.
• Depth or height maps. 2D arrays with a depth value per pixel, plus intrinsics or scale.
• Meshes reconstructed from scans, which may smooth over small defects.
• Companion 2D images captured with the geometry.

The position and orientation of the part relative to the sensor, called the pose, and the calibration that maps sensor units to millimeters, are as important as the points themselves. Without them, labeled defect locations cannot be reproduced.

A reference for multi-view capture only

Real-IAD contains about 150,000 high-resolution images of 30 objects of varied materials, captured from multiple views, and proposes a setting where defective images appear in training (Wang et al., 2024). It is an image dataset, not a point-cloud or profilometry collection. Use it only as a reference for how multi-view capture can be organized per object and how a training setting can tolerate defective samples. It does not tell you what a 3D archive from a specific plant contains.

What labeled 3D data can support

• Geometric defect detection, such as dents, warps, burrs, sink marks, missing features and surface waviness.
• Measurement-linked inspection, if labels are tied to tolerances; see dimensional inspection records.
• Anomaly detection on nominal geometry, with a reference scan or CAD comparison.
• Pose and shape perception, which is a distinct use covered in object recognition versus manipulation data.

Choosing between 3D and 2D for your use

Not every defect needs geometry. If the target is a stain, a color mismatch or a printed-mark fault, a 3D archive adds cost without information. If the target is a depth-related feature such as a dent, a sink mark, a gap between mating parts or a warped edge, 2D images under raking light can approximate it, but a height map states it directly. Ask the supplier whether the same parts were also imaged in 2D, because paired 2D and 3D records let you test which modality your task needs before you commit to a larger purchase.

Also decide how the data will be consumed. A model that takes point clouds directly, one that takes rendered depth images and one that fuses geometry with color need different storage and preprocessing. Fixing the consumption format early keeps you from licensing a large archive in a form you must convert at a loss.

Limits: what this data cannot support

It cannot teach manipulation. A scan of a part shows what it looked like on a fixture or conveyor, not how to grasp it, how it deforms under a gripper or how heavy it is. A humanoid acting on the part needs robot or teleoperation data with synchronized state and action; scans only help it perceive.

It does not transfer automatically between sensors. Resolution, noise, occlusion and reflectivity behavior differ by sensor type and setup. Shiny or dark surfaces produce dropouts. A model trained on one scanner's output can fail on another's.

It lacks context if parts are always presented in one pose on one fixture. A robot seeing parts in bins or in hand faces different views.

Labels may be coarse. A part-level "reject" does not locate the defect. Point-level or region-level labels are costly, and archives usually do not have them unless a vision system required them.

Illustrative package specification

The following is an illustrative specification with synthetic values, not an existing dataset.

# Illustrative only: synthetic values, not a real or available dataset
package: surface_3d_inspection_v0
parts: {family: stamped_bracket_A, lots: 12, sites: [line_B, line_D]}
scans:
  format: [ply, tiff_16bit_heightmap]
  units: millimeters
  sensor: {type: laser_profiler, x_resolution_mm: "recorded per scan", z_resolution_mm: "recorded per scan"}
  calibration: sensor_to_part_transform per scan
labels:
  level: region        # alternatives: part, point
  classes: [dent, warp, burr, missing_feature, none]
  disposition: [accept, rework, scrap]
companion_images: optional, same fixture
splits: by lot and line, never by random scan
exclusions: robot actions, grasp data, forces
rights: training and evaluation; no redistribution

Replace the sensor details with those of your intended deployment. If your robot will use a depth camera, scans from a lab-grade profiler may be useful for definition of defects but not for training on that sensor's noise; see imaging conditions metadata.

Quality and acceptance criteria

You decide the thresholds. Checks that tend to expose weak data: open a sample of scans and verify that labeled regions coincide with a visible geometric feature; confirm that calibration is present and consistent; check that each defect class has examples from more than one lot; confirm that "none" parts were verified, not merely unflagged. Read defect annotation formats before agreeing on label structure, and see metal surface defects for a related material-specific page.

Permissions and provenance

Scans of parts can reveal designs. A 3D point cloud of a customer's component is close to a copy of its geometry. Ask who owns the part design, whether CAD-equivalent geometry is protected by customer contracts, and whether the supplier can license scans of customer parts. Anonymizing part numbers does not remove the design. This is not legal advice.

Questions to ask suppliers

1. Which sensor types and models produced the data, and are raw outputs retained?
2. Is calibration stored per scan, and in what coordinate frame?
3. Are labels at part, region or point level, and who made them?
4. What defect classes appear, and how many lots represent each?
5. Are nominal or reference scans available for comparison?
6. Do scans include dropouts or filtered regions, and is filtering documented?
7. Are the parts the supplier's own designs or customer designs?
8. Can scans be exported in an open format with metadata?
9. Can a small sample be reviewed before a license is agreed?

Discuss your requirements

If you need 3D inspection data, describe the sensor type, part family, labels and permitted uses 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. For engineering-file context, see the catalog entry on CAD and PCB engineering files.

Tell SourceX what your evaluation set needs to cover

Sources

• Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection — multi-view image capture reference, not a 3D dataset.