Imagine that you are designing an inline cell for a customer's foundry to check pump housings against CAD. The camera looks fine on the datasheet. Then the cell runs its first real shift, and the point cloud comes back full of holes wherever a machined bore catches the light. A batch of good housings gets flagged as scrap, your customer calls, and nobody changed the part. The camera just never saw it properly in the first place.
That gap between a spec sheet and a working cell is where most 3D camera selections go wrong. Industrial part inspection covers an enormous range of parts and tolerances: the camera that is right for your customer's stamped brackets is rarely right for their cast valve bodies or turbine blades.
Turbine blades awaiting 3D inspection for dimension correctness.
The largest application is dimensional inspection and metrology: capturing a point cloud of a part, registering it against the reference CAD model, and evaluating it against GD&T callouts such as hole position, flatness, and wall thickness. Done in line and without contact, this is how customers move quality control off a coordinate measuring machine, accurate but slow and offline, and onto the production line itself.
Closely related is surface and cosmetic defect detection, where 3D height data is paired with 2D texture to catch dents, pitting, corrosion, and casting porosity. Weld inspection checks seam geometry and bead height, and assembly verification confirms a bolt or bracket is present and correctly seated before the next station. Sorting, classification, and reverse engineering a worn part back into CAD round out the list. Each depends on the same foundation: a complete, trustworthy point cloud.
Most 3D vision techniques assume light scatters diffusely off a surface. Bare, machined, or polished metal instead reflects specularly, like a mirror, and that breaks several things at once. Highlights saturate the sensor, and light that glances off a curved surface away from the sensor never returns at all, leaving dropout, literal holes in the point cloud. Worse, inside a hole or a tight corner, light can bounce more than once before reaching the sensor, and structured light systems can misread that interreflection as a false depth reading that bulges or warps the reconstructed geometry rather than just adding noise.
Complex metal parts such as this gearbox housing present significant high-contrast challenges.
Dark, oxidized, or sandblasted metal has the opposite problem, absorbing light and starving the sensor of signal. Parts fresh off a machining line often carry coolant, oil, or swarf, adding another reflective layer. Add ambient light that shifts through the day, glare from welding arcs, dust, vibration, and temperature swings that drift a camera's calibration, and it is clear why a camera that performs well in a lab can struggle on a production floor, against tolerances measured in tens of microns and a cycle time that has to keep pace with the line. One thing works in your favor: inspection rarely involves a robot gripping the part, so you are spared the collision risk that comes with bin picking. The risk that remains is simpler but just as costly: an image that is incomplete or inaccurate.
Trueness is the number to anchor your spec on, not just repeatability, since a camera can be highly repeatable and still consistently wrong. Zivid 3D cameras typically hold trueness error below 0.2 percent, which on a 30-centimeter part works out to a deviation from reality of around 0.06 millimeters or less, the margin that decides whether good parts pass and bad parts get caught.
Technical terms: learn more about trueness error and the difference between accuracy and precision →
High dynamic range matters because metallic parts are often high in contrast, especially with oil or grease present, and a camera without strong HDR will blow out the highlights or lose the dark recesses in the same capture. Point cloud completeness is arguably the most inspection-specific requirement of all, since a feature that never made it into the cloud cannot be measured. Specialist vision engines tuned for reflective metal and glossy plastic suppress the false readings that interreflection normally triggers, rather than just averaging through the noise.
Field-of-view has to match the part, not just the camera's headline resolution: getting closer improves accuracy, but the working distance still needs to cover the part in as few captures as possible. Capture speed matters too, since a scan that completes well under a second keeps the vision system from becoming the bottleneck. And because this equipment lives on the shop floor, industrial robustness, tolerance for dust, vibration, and temperature swings without drifting out of calibration, is not a nice-to-have. A camera that drifts is worse than no camera at all: it produces confident, wrong measurements.
Both mounting approaches show up in the field. The common one puts the part on a rotating platform, capturing, rotating, and stitching the results into a full 360-degree model. The alternative mounts the camera on a robot and moves it to multiple viewpoints and distances, letting you close in on a feature that needs extra resolution.
A mobile inspection robot system with a robot-mounted 3D camera inspecting locomotive wheels.
It is worth being just as clear with your customer about fit as about capability. 3D inspection is at its strongest on medium to large parts where the question is correct form and manufacture: chassis structures, machined housings, castings, turbine blades, brackets, and connectors. It is not the right tool for finding hairline cracks or fine scratches on glass, or for parts moving continuously on a conveyor, since that level of fidelity sits below what any 3D camera can reliably guarantee. Setting that expectation early saves a difficult conversation later.
Mounting your camera on the robot: why it is a deal-changer for your application →
Zivid recommends the Zivid 2+ MR60 as the starting point for most high-performance part inspection, built for both static and on-robot mounting, with unified 2D and 3D data, including surface normals and per-pixel signal-to-noise ratio, so inspection algorithms and AI models alike get clean input. Infield Correction lets you verify and restore trueness out on the floor rather than shipping a camera back to the factory, and native Hand-Eye Calibration handles the robot-mounted case without a separate routine of your own. Pandia 3D built its Pure inspection platform on the same logic, using Zivid 3D cameras to let manufacturers scan parts and check them against CAD without specialist expertise or an enclosure.
Specifying a 3D camera for part inspection comes down to matching trueness, dynamic range, and completeness to the material and tolerance the part demands, not chasing the best headline resolution. Get that match right, and the cell you deliver becomes the fast, trustworthy replacement for the offline CMM queue it was meant to be.
Would you like to learn more about inspection applications with Zivid? Schedule a meeting with our engineers and get a demo suited to your needs: