Coplanarity, protrusion and seal gap are height measurements, and a 2D camera cannot measure height — it sees a flat projection, so it can guess these from shadow or focus but it cannot gauge them. When a 2D system fails on these features it is the wrong instrument, not a wrong model, and the fix is a sensor that measures the third dimension, not more training data for the one that cannot.
That is the leading-indicator point. A repeated failure on lead coplanarity or a seal-gap check is the earliest signal that the measurement method is mismatched to the feature, long before anyone blames the AI. Height is a metrology problem. You solve it by choosing an instrument that resolves height, then letting a model work on data that actually contains the dimension.
The feature-to-method map
| Measurement | What it is | Method that measures it | Key spec that decides fit |
|---|---|---|---|
| Coplanarity | Height variation of leads/balls to a seating plane | Laser triangulation; structured light | Z-resolution vs tolerance |
| Protrusion | How far a feature stands proud | Laser triangulation; stereo | Standoff and occlusion |
| Seal gap | Height/width of a gap along a seam | Structured light; laser profile | Lateral + Z resolution together |
Three 3D methods, three trade-offs:
- Laser triangulation — high Z-resolution on a scanned line; needs relative motion and can be occluded by tall neighbours. Strong for coplanarity and protrusion.
- Structured light — a full-field height map in one capture; excellent for seal gaps and surfaces, sensitive to shiny or very dark surfaces.
- Stereo — two cameras infer depth; flexible and lower-resolution, good where tolerances are looser and the surface has texture.
The specs that decide the choice are Z-resolution against your tolerance, standoff, and cycle time — a method that resolves 10µm but cannot keep the line's 0.3-to-1.0-second-per-unit pace is still the wrong instrument for that line.
Why "add more images" is the wrong reflex
When a 2D check keeps missing a coplanarity reject, the instinct is to label more images and retrain. It cannot work, and the reason is physical, not statistical: the height information was never in the 2D image to begin with. A flat projection does not encode how far a lead sits off the seating plane, so no quantity of flat images teaches the model a dimension the camera did not capture.
This is the same error as blaming the model for a wavelength problem: the discipline is to ask whether the feature is even present in the data before touching the training set. For height features on a 2D sensor, it is not. Change the instrument first.
Where 2D is still right, and the honest boundary
None of this is an argument that 3D is always better. A 2D camera is the right, cheaper instrument for in-plane features — presence, position, print, surface defects, dimensions in the X-Y plane — and over-specifying 3D where 2D suffices is its own waste, the mirror of the error above. On one deployment a 2D solution met a standard where 3D had been proposed, at roughly a third of the cost (Client B, public alias), precisely because the features were in-plane. The rule is to match the instrument to the dimension of the feature, in both directions.
The honest boundary: the feature-to-method table is metrology reasoning, and the resolution and cycle-time figures (10µm, 0.3 to 1.0 seconds per unit) are platform specs on suited tasks, not a claim that every 3D feature is solved. The test for your part is a sample measured against your tolerance on the candidate method.
Related reading: the field-of-view and resolution math post, the gauge R&R for AI inspection post, the colour, gloss and texture metrology versus AI post, and /solutions/hyperq-ai-vision.
Frequently asked questions
Why can't a 2D camera measure coplanarity? Because coplanarity is a height variation and a 2D camera records a flat projection with no true height information. It can infer from shadow or focus but cannot gauge the dimension.
Will more training data fix a 2D height-measurement failure? No. The height was never captured in the 2D image, so there is no signal to learn. The fix is a 3D method — laser triangulation, structured light or stereo — not more images.
When is 2D still the right choice? For in-plane features: presence, position, print and X-Y dimensions. Specifying 3D where 2D suffices wastes money, just as using 2D for a height feature fails the measurement.
Send one part with the coplanarity, protrusion or seal-gap tolerance you need to hold, and within two weeks we will return which 3D method resolves it at your line speed, or confirm 2D is enough, before any commitment. Start at https://apac.hypernology.net/contact.