3D Vision in Manufacturing: Stereo, Structured Light, and Where Depth Earns Its Keep

3D vision in manufacturing earns its keep wherever the defect or the task is fundamentally about height, shape, or position in space: bin picking, coplanarity checks, weld bead and adhesive bead inspection, flatness, and volume measurement. Four technologies dominate: laser triangulation, structured light, stereo vision, and time-of-flight, and they differ by orders of magnitude in resolution, speed, and price. The honest starting point is that 3D is more expensive and more complicated than 2D, so the first question is always whether a well-lit 2D image would do the job.

Here is how the technologies compare, where each fits, and the practical problems the brochures skip.

When 2D isn't enough

A 2D camera flattens the world. Contrast is all it has, so anything that changes gray levels, such as print, stains, holes, and edges, is fair game, and good lighting can manufacture contrast for a lot of surface features. But some tasks defeat contrast entirely.

A dent in an unpainted metal panel may produce almost no gray-level change while being an instant customer rejection. A connector pin bent 0.3 mm out of plane looks identical from above. A dollop of clear adhesive is nearly invisible to a camera but has a height and a volume that determine whether the joint holds. And a robot reaching into a bin of jumbled castings needs to know where the top part is in millimeters, not pixels.

The pattern: when the specification is written in height, depth, volume, flatness, or 3D position, a 2D image is answering the wrong question. That is the boundary where depth sensing starts paying.

The four main 3D vision technologies

Laser triangulation projects a laser line onto the part and views it with a camera at an angle; height changes shift the line in the image, and simple geometry converts that shift to height. One image gives one profile, so the part or the sensor must move, with an encoder tying profiles together into a full scan. Modern laser profilers capture thousands of profiles per second at height resolutions from a few microns to tens of microns, which makes them the workhorse for inline scanning of anything already moving: extrusions, weld beads, adhesive paths, tire treads, brake pads, machined faces. Typical sensor prices run $5,000 to $30,000 depending on resolution and speed.

Structured light projects a sequence of patterns, usually fringe patterns, from a projector while a camera watches how the surface warps them, recovering a full-field 3D snapshot without motion. Resolution reaches microns to tens of microns over small to mid-size fields, and acquisition takes a fraction of a second. It shines for stationary measurements: electronics inspection, solder paste and component coplanarity, dimensional checks of molded and machined parts, and robot guidance where the scene holds still for a moment. Industrial snapshot sensors typically run $10,000 to $50,000.

Stereo vision uses two cameras and matches features between views, like human depth perception. Passive stereo needs surface texture to match against and struggles on plain surfaces, so most industrial stereo is active, adding a projected random pattern to guarantee texture. Stereo trades some precision, typically sub-millimeter to a few millimeters, for large fields of view, speed, and moderate cost, which suits logistics, depalletizing, and bin picking of larger objects more than precision gauging.

Time-of-flight measures how long emitted light takes to return, per pixel, producing depth maps at video rates over long ranges. Resolution is coarse, typically several millimeters, but nothing else covers room-scale volumes at 30 frames per second for a few thousand dollars. ToF belongs in pallet dimensioning, fill-level checks on large containers, vehicle and object profiling, and safety and presence sensing, not in inspection tolerances.

Which 3D vision technology fits your application?

Three questions sort nearly every application.

First, is the part moving or stationary? Continuous motion with an encoder is laser triangulation territory; the motion you already have becomes the scan axis for free. Stationary parts, or robot-presented parts, favor structured light snapshots.

Second, what resolution does the tolerance demand? Work backward from the feature: verifying a 50 micron coplanarity spec needs micron-class structured light or a high-end profiler, while locating castings in a bin is comfortable at a millimeter. Buying ten times the resolution you need buys you ten times the data and none of the benefit.

Third, how big is the scene? Micron resolution and meter-scale fields do not coexist in one affordable sensor. Large scenes at moderate precision point to stereo or ToF; small fields at high precision point to structured light or triangulation. When someone asks for micron precision across a full pallet, the real answer is a different system architecture, perhaps a profiler on a motion axis, or a re-scoped requirement.

Cost tracks those answers. As rough planning figures: ToF cameras typically $500 to $5,000, active stereo systems $2,000 to $15,000, laser profilers $5,000 to $30,000, precision structured-light sensors $10,000 to $50,000, before any of the integration work that dominates system cost in 3D just as it does in 2D.

Where depth earns its keep on real lines

Bin picking is the marquee application: a 3D sensor above a bin finds part poses, collision-checks a grasp, and hands coordinates to a robot. It has matured from research demo to dependable production tool, though success still depends on gripper design and what fraction of the bin the system can actually pick, a number worth pinning down in acceptance criteria rather than assuming to be 100 percent. This is as much a robotics and automation problem as a vision one.

Bead inspection, for welds, sealants, and adhesives, is quietly one of the highest-value uses. A profiler scanning a bead verifies height, width, and cross-sectional area continuously, catching gaps, thin sections, and air bubbles that a top-down camera cannot judge. On structural adhesives in particular, volume is the specification, and only 3D measures it.

Electronics manufacturing leans on 3D for solder paste inspection and component coplanarity, where everything interesting happens in tens of microns of height. Assembly verification uses depth to confirm a clip is fully seated or a gasket sits proud by the right amount, checks that are ambiguous in 2D and trivial in a height map. And at the coarse end, ToF and stereo handle pallet dimensioning, truck fill, and case counting in warehouses without any precision pretensions.

The practical gotchas

3D sensing has failure modes that surprise teams coming from 2D. Occlusion is first: triangulation-based methods need both the projector and camera to see a point, so tall features cast data shadows, and steep walls vanish. Dual-head profilers and multi-view scanning fix it, at a price.

Surface optics is second. Shiny and specular parts throw laser lines into multiple reflections that produce phantom spikes; translucent plastics let the laser penetrate and blur; black rubber absorbs so much light that the profile starves. HDR scanning modes, wavelength choice, and exposure tuning recover most of these, but budget bench time with your real parts, because a demo on a machined aluminum block proves nothing about your glossy black molding.

Data volume is third. A profiler producing 2,000 profiles per second at 2,000 points each is generating four million 3D points per second, which is why 3D systems ship with 10 GigE interfaces and why point cloud processing wants serious CPU or GPU behind it. Filtering, meshing, and alignment steps that felt instant on a sample scan can quietly blow a cycle-time budget. And calibration is a discipline of its own: the sensor-to-encoder and sensor-to-robot transforms drift with temperature and collisions, so plan verification artifacts and a recalibration routine, not a one-time setup. Deep learning is also arriving in point cloud interpretation, and the same rules-versus-learning tradeoffs apply in 3D as in 2D.

FAQ

What accuracy can 3D vision achieve in manufacturing?

It spans four orders of magnitude by technology: precision structured light and high-end laser profilers resolve single microns of height over small fields, mainstream profilers tens of microns, active stereo sub-millimeter to a few millimeters, and time-of-flight several millimeters. Match the technology to the tolerance, and be skeptical of accuracy claims quoted without a field of view.

Is 3D vision better than 2D vision?

Neither is better; they answer different questions. 2D reads contrast, so print, presence, and surface appearance belong to it, at lower cost. 3D measures shape, so height, flatness, volume, and spatial position belong to it. Many strong systems combine both, using a 2D image for surface checks and a height map for geometry in one station.

What does a 3D vision system cost?

Sensors alone typically run from about $500 for time-of-flight cameras to $50,000 for precision structured-light units, with laser profilers commonly $5,000 to $30,000. Complete integrated systems, including handling, compute, and controls integration, typically land between $50,000 and $200,000 or more, with integration effort claiming a share of budget similar to 2D projects.

Why do shiny or black parts cause problems for 3D scanners?

Triangulation methods depend on the surface returning projected light diffusely toward the camera. Mirror-like surfaces reflect it off-axis or create false multi-bounce returns, and very dark surfaces absorb too much for a clean signal. HDR scan modes, coating sprays for offline work, exposure strategy, and sensor selection all help, which is why testing on your actual parts comes before purchasing.

If your specification talks in microns of flatness or millimeters of robot reach, depth sensing is probably the right tool, and choosing among these four technologies is the decision that sets your budget. Willowark engineers vision and sensing systems in both 2D and 3D and can test your parts before you commit to an approach. Reach out and bring your hardest sample.