3D vs 2D machine vision system: when to choose 3D

Every inspection asks one of two questions: what does the part look like, or where is its surface in space. A 2D camera answers the first — brightness per pixel, in milliseconds. A 3D machine vision system answers the second — height, flatness, volume and pose, whatever the color or the finish does. The choice follows from the sentence your inspection report has to produce, not from the budget line. This comparison gives you the four-question test that decides it, the numbers behind the decision (pixel size, height repeatability, tolerance ratio, scan time), a worked bead-inspection example with the arithmetic shown, and the cases where the popular answer is wrong.

What does a 3D machine vision system measure that a 2D camera cannot?

A 3D machine vision system outputs geometry — height, flatness, warp, volume and 6-DOF pose — instead of brightness, so a 0.2 mm warp stays measurable when color, sheen or print change from part to part; a 2D camera only reports how much light each pixel received.

A 2D image is an intensity map. Height reaches it only sideways — as a shadow, a focus loss, a shading gradient — and every one of those channels also moves when the finish, gloss or print moves. That is why a 2D recipe tuned on the first batch drifts as tooling wears. A 3D sensor projects a known geometry (a laser line, a coded pattern) or matches two views, then computes distance per point. The classic 3D jobs follow directly: height steps, flatness over an area, hole depth and position, dent and volume checks, and full object volume — all finish-independent. To see how one such measurement reads out, measuring surface height difference from a 3D point cloud walks through a real step.

The 3D sensors category splits into laser profilers, structured-light stereo, dome-based 2.5D systems and time-of-flight cameras — the four technologies compared below.

Which inspection jobs still belong to a 2D camera?

Presence and absence, label position, date codes, barcodes, color and one-dimensional gauging belong to 2D: one exposure of 0.1-10 ms answers them, with none of the scanning, encoding or point-cloud processing a 3D sensor adds.

Lateral precision is not the limitation — sub-pixel edge fitting reaches about a tenth of a pixel, 2.5 µm in the worked example below. A vision sensor or smart camera runs these checks on the device itself: the MV-SC5060XM-08M-WBN pairs a 6 MP mono sensor with an 8 mm lens and needs no PC. Barcodes and print also survive dirt better in 2D than in a point cloud; 1D barcode preprocessing for rotation, blur and glare covers the failure modes you will meet.

For the camera itself, how to choose a vision inspection camera covers sensor size and interface; for making the station stable, stable 2D vision inspection systems puts lighting, lens and exposure in the right order. Browse area scan cameras or smart cameras when the job is contrast, and the camera finder filters by interface and shutter.

When does my inspection need 3D rather than 2D?

Choose 3D when any one of four is true: the output is a height, flatness or volume; the part arrives in random pose or stacked; the defect vanishes under every lighting you tried; or color and finish vary part to part.

One yes sends you to 3D. Zero yes answers mean a 2D camera with the right lighting is the faster, simpler system, and your money belongs in the lighting. Use the table at the quoting stage:

Random pose and stacks are the robotics cases: 2D or 3D camera for conveyor pick and place works that decision for moving conveyors, and 3D bin picking for flat parts handles parts that look identical from above. When all four answers are no, stop before 3D: metal surfaces are the classic case where lighting wins first — lighting for metal surface defect inspection beats glare long before a laser line is needed.

Which 3D technology fits my part?

Use a laser profiler on moving parts, projected structured light on stationary parts scanned in one shot, a dome-based 2.5D station on glossy or wrinkled surfaces, and time-of-flight only when a coarse depth map at meters of range is enough.

Laser triangulation throws a laser line across the part and watches it with a camera at an angle; height comes from where the line lands on the sensor, and an encoder stitches profiles together as the part moves. It is the default for inline height and flatness checks, holding micron-class repeatability at short standoff. It needs a surface that scatters — not a mirror — and tall walls near the feature shadow the line. The MV-DP4180-03P is a 3D laser profile sensor of this class.

Projected structured light paints coded patterns on a stationary part and decodes depth per pixel from how the pattern deforms — one shot, full field, no encoder. It dislikes ambient light and any part motion during the pattern sequence. Stereo pairs with structured light add stability on larger scenes: the MV-DLS300P-04 laser galvanometer stereo camera and the MV-DPS200P-02 projected structured-light 3D camera cover the two variants.

Dome-based 2.5D stations light the part from every direction at once and read height from shading — what glossy, curved and wrinkled surfaces need, since a laser line scatters into streaks on a mirror. The MV-PSA-12UM-C4D30-WDM is one such system with a 12 MP camera and a 300 mm dome light.

Time-of-flight cameras measure per-pixel return time and give a whole-frame depth map at meter range with centimeter-class accuracy — right for volume and presence over large scenes, wrong for a 0.2 mm flatness call. The MV-DT01WMU ToF sensing camera is the format example; before you pick any ToF unit, read choosing a time-of-flight unambiguous range, because the modulation frequency sets a hard maximum distance.

What resolution does a height check need?

Size the sensor from the tolerance, not the field of view: keep height repeatability at or below a quarter of the tolerance band (a 4:1 ratio, 10:1 comfortable), and lateral resolution at about a tenth of the smallest feature you must see.

Worked example — black EPDM bead on black PP housing. Part 100 mm × 80 mm, bead height spec 0.8 mm ± 0.2 mm, bead width 2 mm, one part every 4 s.

The 2D attempt: a 1/1.7"-class sensor is about 7.4 mm wide with 4000 pixels across, so each pixel is 7.4/4000 ≈ 1.85 µm. Over a 100 mm FOV one pixel covers 100/4000 = 0.025 mm, and the 2 mm bead spans 2/0.025 = 80 pixels — width and position are easy. Magnification is m = 7.4/100 = 0.074, so a 25 mm lens sits at WD = 25 × 1.074/0.074 ≈ 363 mm. Depth of field at f/8 (two-pixel circle, estimate) is about ±6 mm, plenty for a located part. Exposure: stopped, milliseconds are enough; if the part runs past at 100 mm/s, keeping blur under one pixel (25 µm) caps exposure at 0.025 mm ÷ 100 mm/s = 0.25 ms, strobed. All of that works — and none of it measures height: black on black gives a few gray levels under any flat lighting, and the 0.8 mm the spec asks for is not in the image.

The 3D answer: a mid-range laser profiler at this standoff delivers height repeatability near 0.03 mm (typical for the class; short-standoff units reach single-digit microns). The tolerance band is 2 × 0.2 = 0.4 mm, so TAR = 0.4/0.03 ≈ 13 — above 4 passes, above 10 is comfortable. Scanning: 80 mm of travel at 0.2 mm profile pitch needs 80/0.2 = 400 profiles; at 2000 profiles/s that is 0.2 s inside the 4 s cycle. Decision: laser profiler for the height, one 2D camera for the label.

These are example inputs for a mid-range class, not one catalog number — short-standoff profilers do several times better, long-range units worse. For your own field of view and feature size, run the same arithmetic here:

The 2D half of the station can be modest: the LS-CMR-0166 is a 12 MP USB3 camera with a 1/1.7" sensor and a 4000 × 3000 resolution, one frame per part.

What does 3D cost in cycle time and integration?

Plan for 0.1-2 s of scanning, an encoder or trigger per profile, and a point-cloud pipeline your team builds and calibrates — roughly three to five times the engineering of a 2D station that answers with one frame.

The line-speed ceiling follows from pitch: 0.2 mm profile pitch at 2000 profiles/s tops out at 0.2 × 2000 = 400 mm/s of conveyor speed. Faster lines need faster profilers or line-scan 2.5D — the MV-HLVS-16K100-L60 scans at a 120 kHz sampling rate with a 600 mm emitting area, sized for full-speed production.

Integration is the second cost. A profiler needs an encoder input and clean triggering; a stereo projector needs a still part; the cloud needs calibration, filtering and measurement code on your side. Two settings decide your effective precision: depth unit and precision settings on 3D cameras explains what they change, and the 3D camera working distance and positioning guide covers the standoff window you must mount inside. Data volume is fair: 400 profiles × 1000 points × 3 axes × 4 bytes ≈ 4.8 MB per part against 12 MB for a 12 MP mono frame — comparable to carry, heavier to process.

When is 3D the wrong answer even for a height check?

Skip 3D when the height difference already casts a shadow a 2D camera can count on, when the surface mirrors along the triangulation plane, when the part is clear glass, or when the line has no seconds to give to a scan.

The shadow limit: a 0.5 mm step lit at 20° above the horizon throws a shadow 0.5/tan(20°) = 1.4 mm long, about 56 pixels at 25 µm per pixel. If the spec only asks whether the step exists, grazing dark-field light and a 2D camera answer in one frame; if it asks for the number, they cannot. Photometric stereo goes one step further — photometric stereo ring lighting for surface inspection recovers surface slope from four lighting directions, enough to grade texture and dents without a point cloud.

Mirrors, glass and clear plastics defeat triangulation outright: the line passes through or reflects away, and the cloud comes back with holes. Wrinkled, glossy surfaces belong on a dome 2.5D station instead. And when the tolerance is ±0.5 mm or looser on a fixtured part, a shadow check or a mechanical gauge answers for a fraction of the engineering.

What should I check before ordering a 3D sensor?

Five checks: repeatability on your material at your standoff, scan rate against line speed, the working-distance window against mounting space, interface and power against your cabinet, and environment — heat, vibration, washdown — against the sensor's rating.

• Height repeatability measured on your material and finish, at your standoff — a sample run beats any brochure line.
• Scan rate against conveyor speed: pitch × profiles per second must exceed line speed with margin.
• Working-distance window and mounting envelope, including the triangulation blind side near tall walls.
• Interface and power you can actually supply: encoder input, trigger line, PoE or 24 V.
• Environment: heat, vibration and washdown against the housing rating; if the station sits in a washdown zone, the enclosure and its glands carry that rating — NEMA 4X versus IP66 ratings explains what each claim covers.
• Laser class and station labeling for the people who run the line.

Run the four-question test on your worst-case part, then size the tolerance ratio and the scan time as in the example above. When the answer is geometry, start among the 3D sensors below; when it is contrast, the 2D finders linked above do the job. You can also step through the system configurator, and the 3D topic page collects the point-cloud measurement and set-up guides. Free expert support comes with every order, including help reading samples of your part before you commit.