3D versus 2D Inspection: Where the Boundary Lies

The question is not which technology is better. It is whether the judgement can be expressed as a projection. A 2D camera records the light arriving along one direction, so two scenes that project to the same pattern are indistinguishable to it. A part resting correctly on a seat and a part hovering three millimetres above that seat can produce exactly the same image if the geometry lines up, and no amount of resolution resolves it, because the information was never captured. Once that is understood, most 3D questions answer themselves.

The five things 2D cannot judge

• Height and seating. Whether a component is fully pressed home, seated on its boss, or resting on a swarf particle. Visually identical to a correctly seated part from directly above.
• Gap and flushness. The gap between two panels, or how far one surface stands proud of its neighbour. These are differences along the viewing direction, so they are exactly what a projective camera discards.
• Coplanarity. Whether three or more contact points lie in one plane. Commonly checked by pressing on a surface and looking for rock, which is a mechanical proxy precisely because vision cannot see it.
• Volume and presence of material. The size of an adhesive bead, the fill level of a dispensed dome, the amount of grease in a joint — the outline may be identical while the quantity is not.
• Depth-ordered occlusion. Whether one part is in front of another or merely projected over it. A 2D system must be told what is in front; a 3D system can determine it.

Four methods, and what each one costs

The cost side is often understated. A 3D system carries a heavier calibration, because a depth sensor has its own intrinsic and extrinsic parameters on top of the usual lens model, and its accuracy varies across the field in a way a 2D camera does not. Data volume rises by an order of magnitude, which puts pressure on bandwidth and compute exactly where cycle time is tight. And the surfaces that are hardest for 2D are frequently hardest for 3D as well: a shiny machined face that blows out a 2D image also fails to return a reliable laser line.

The boundary rule, and what to do at it

Write the judgement as a sentence. If every term in that sentence can be measured in the image plane — length, width, area, position, angle within the plane, presence of a feature — then 2D is sufficient, and a well-designed 2D station will be cheaper, faster and easier to keep in calibration. If any term requires distance along the viewing direction, then 2D cannot make the judgement at all and no improvement to the camera will change that.

Right at the boundary, the practical alternatives are worth pricing before reaching for 3D. A second camera viewing from the side is sometimes enough to capture the one dimension that matters and costs far less than a full depth sensor, because it converts a depth question into a lateral measurement in the second view. A mechanical gauge plus a vision presence check is another: the gauge establishes the height, the camera confirms the gauge was offered and the part was present. Both are real options, and both are considerably cheaper than a full 3D station. What should decide the choice is the wording of the judgement, not the technology in the catalogue. The same discipline applies to the decisions that should not be automated at all, which is the subject of the boundary of motion criteria.