What is a 3D Camera? Complete Guide for Computer Vision

What a 3D camera is

A 3D camera in Computer Vision is a sensing system that measures depth, distance from the sensor to points on the object, so software can work with shape, height, volume or pose, not only with a flat intensity image. The output is often a point cloud, a depth map, a height profile, or a combination of depth plus a 2D intensity or color image.

That is a different question from area scan or line scan. A 2D camera asks “what does it look like?” A 3D system asks “what is its geometry in space?” If your defect is a missing label, wrong print color, or a scratch that shows in contrast, stay in 2D. If you need bin picking pose, fill volume, warpage, connector coplanarity, or glue-bead height, 3D enters the conversation.

“3D camera” is an umbrella term, not one sensor physics. Under it you will find stereo vision, structured light, laser triangulation profilers, time-of-flight variants, and RGB-D devices that fuse color with depth. In Sedeco’s taxonomy they sit under Cameras as a family alongside area scan, line scan and smart cameras. The goal is to help you choose a technology class and ask the right feasibility questions.

How industrial 3D capture works (buyer view)

Ignore brand marketing for a moment. Most factory 3D systems share four practical stages: illuminate or observe geometry → compute depth → deliver a cloud or profile → decide in software.

1. The scene must reveal depth

Depth measurement needs a physical cue: disparity between two views (stereo), deformation of a projected pattern (structured light), the shift of a laser line on a sensor (triangulation profiler), or timed light return (ToF-style approaches). Shiny, transparent, or very dark surfaces can starve those cues. Material finish is therefore a first feasibility question, not a footnote.

2. Depth is computed, not photographed

Unlike a simple intensity frame, 3D data is calculated. Algorithms match features, fit laser lines, or decode patterns. That computation has latency, working-distance limits, and accuracy that depend on calibration, baseline, optics and temperature stability. Datasheet “Z accuracy” is always conditional, on distance, surface, and setup.

3. Output formats differ by technology

A laser profiler often delivers a height profile per scan line, built into a surface as the part moves, conceptually related to line scan motion. A stereo or structured-light snapshot may deliver a dense depth map of a volume in one shot. RGB-D devices add a registered color image for texture or ID alongside depth. Match the output to your software stack before you fall in love with a brochure render.

4. Software decides pass/fail or pose

Point clouds and depth maps are raw material. Robot guidance, volume calculation, CAD comparison and bead inspection live in software, on a PC, a vision controller, or sometimes onboard a smart 3D device. Architecture still matters; see smart camera vs PC-based vision when processing location is undecided.

3D vs area scan, line scan, and smart

Do not buy 3D because it sounds advanced. Buy it when the decision needs geometry.

3D vs area scan: start with the area scan guide if a well-lit 2D view can show the defect. Many “height” problems are actually contrast problems solved with better lighting, see lighting.

3D vs line scan: continuous surface print still prefers line scan. Continuous height (for example bead or weld profile along travel) often uses a laser profiler synchronized to motion, a 3D cousin of the line-scan idea.

Where 3D cameras are used

Patterns that commonly justify 3D for project teams:

• Bin picking and robot guidance, Find randomly oriented parts in a bin or on a belt; send pose to the robot.
• Volume and fill, Parcel dimensioning, bag fill, dough or powder height where mass is inferred from shape.
• Height and gap gauging, Connector pin height, gasket stand-off, gap to a datum plane.
• Bead, weld and seal profile, Cross-section height and width along a path.
• Warpage and flatness, Boards, plates and molded parts against a plane tolerance.
• Completeness when contrast fails, Dark-on-dark clips or transparent features that disappear in 2D but change geometry.

If the feature is purely cosmetic print or a high-contrast edge in a controlled light, 3D is usually the wrong spend. If operators already struggle to “see height” by eye, 3D may be the honest tool.

Specs that actually matter when you buy

3D datasheets mix accuracy, FOV, point density and frame or scan rate. Prioritize the list below.

Working distance and field of view

Every 3D technology has a usable volume, a box in space where accuracy holds. Parts outside that volume degrade or vanish. Measure your station envelope and part placement variation before comparing Z numbers.

Z accuracy and repeatability

Accuracy is how close measured height is to truth; repeatability is how stable repeated measures are. teams often need repeatability more than absolute accuracy for pass/fail. Ask under which surface, distance and temperature the number was measured.

Lateral resolution / point density

Fine features need enough points across the feature, just as 2D needs pixels on feature. Sparse clouds can miss thin beads or small gaps. Density also drives compute cost.

Acquisition time vs takt

Snapshot 3D (stereo / structured light) has exposure and compute time per scene. Profilers need motion and a scan duration. Confirm the full pipeline fits takt with margin, including robot clearances and PLC handshake.

Surface and ambient sensitivity

Specular metal, black rubber, clear plastic and outdoor ambient IR can break naive demos. Plan sample trials. Optics and sometimes polarization or coating tricks help, lens and light choices still matter; see lenses.

Interface and software ecosystem

GigE, USB, proprietary links and vendor SDKs vary. Confirm that your robot brand, PLC and vision library can consume the cloud or profile format without a science project.

Building a working 3D setup

A reliable 3D station is a system:

1. Sensor head, technology class matched to FOV, Z need and surface.
2. Optics / laser class, eye safety, baseline, magnification as applicable.
3. Mechanics, stiff mount; vibration kills Z repeatability.
4. Motion (if profiler), encoder or known speed for correct spatial sampling along travel.
5. Calibration routine, plate or artifact procedure operators can repeat.
6. Compute & software, cloud processing, CAD match, robot transform, recipes.

Common first-system mistakes

• Buying 3D for a contrast problem that lighting would fix in 2D.
• Ignoring shiny samples until FAT, then discovering dropouts.
• Underestimating compute time on dense clouds at takt.
• Mounting on a vibrating robot pedestal and blaming the sensor.
• Skipping robot–vision hand–eye calibration time in the project plan.
• Assuming one technology fits all parts in a mixed bin.

How to read a 3D product page

On Sedeco product pages for 3D families you will typically see technology cues, working distance / FOV ranges, and interface notes. Read them against the eight questions:

• Technology class, stereo, profiler, structured light, RGB-D, etc.
• Measurement range / FOV, usable volume in the station.
• Z performance, accuracy/repeatability under stated conditions (ask for conditions).
• Scan or frame rate, ceiling for your motion and resolution mode.
• Data interface & SDK, how clouds or profiles reach your software.

When you evaluate a 3D camera, use this guide to understand the category before comparing models in the camera catalogue. Specific millimetre claims belong to validated setups on your parts, not to generic brochure copying.

Stereo, profiler and RGB-D in practical terms

Stereo vision

Two (or more) views of the same scene produce disparity, which becomes depth. Strengths: relatively flexible FOV, often good for robot guidance on textured parts. Limits: texture-poor or specular surfaces reduce matches; baseline and calibration matter. Ambient light and working distance affect performance.

Laser triangulation profilers

A laser line projects onto the part; a camera views the line from an angle; height is inferred from line position. Strengths: strong Z resolution for many gauging and bead tasks; natural fit to conveyors when synchronized to motion. Limits: eye-safety class, specular bounce, and the need for relative motion or scanning to cover an area.

Structured light (snapshot pattern)

A known pattern is projected; deformation yields a dense depth map of a volume in one or few frames. Strengths: full-field height in a station FOV. Limits: ambient light competition, cycle time for projection sequences, and challenging materials.

RGB-D

Devices that deliver registered color (or intensity) plus depth. Useful when you need both texture/ID cues and coarse geometry, logistics, some pick applications, operator visualization. Confirm industrial robustness, working distance and Z precision against your tolerance; consumer-grade RGB-D is not automatically factory-grade.

Industry snapshots

Automotive and metal assembly

Gap & flush cues, weld/bead profiles, clip seating height, and robot guidance on racks. Harsh lighting and oil films push you toward careful technology trials.

Electronics

Connector coplanarity, solder paste / lead height cues, and board warpage. Small Z tolerances demand profiler-class thinking more often than coarse RGB-D.

Packaging and logistics

Parcel cubing, depalletizing guidance, and fill height. Throughput and conveyor motion dominate the design more than micrometre Z.

Food and pharma (selected)

Portion height, seal geometry cues, and pack completeness when 2D fails. Hygiene, washdown and validation constraints shape housing and software more than sensor physics alone.

Integrating with PLC, robot, and IT

Typical loops:

• Trigger, PLC or robot tells the 3D system when the scene is ready.
• Acquire, depth capture; optional motion for profilers.
• Compute, cloud processing, CAD match, pose estimation.
• Act, robot offset, reject, or MES log.

Budget time for coordinate frames: sensor → robot base → part. Many first projects underestimate hand–eye calibration and fixture redesign. Settle PC vs smart processing early via the smart vs PC page.

Cost, quality, and what good enough means

3D is often more expensive than a comparable 2D station, not only the head, but mounts, calibration artifacts, compute and engineering hours. The cheapest failure mode is buying 3D for a problem lighting would have solved.

Define success with samples: golden parts, borderline height fails, and worst-case finishes. If depth dropouts appear on bad finishes in the lab, they will appear on the line. A mid-range profiler with rock-solid mechanics routinely beats a flashy dense-cloud demo on a shaky stand.

When you budget, separate capital cost from project risk. A cheaper head that needs months of experimentation on specular parts can cost more than a clearer technology fit validated in a short feasibility pass. Include operator training and calibration artifacts in the estimate, 3D stations fail quietly when nobody owns the weekly health check.

Good enough means your geometric tolerance is met with margin under real ambient light and takt, not that the point cloud looks impressive in a dark demo booth.

Practical project checklist for first 3D projects

Use this sequence before you freeze a purchase order. It keeps conversations with suppliers factual and reduces late surprises at FAT.

1. Write the geometric decision in one sentence, for example “measure bead height within ±0.2 mm” or “compute pick pose for unsorted brackets.” If you cannot write it, you are not ready for a datasheet.
2. Collect representative samples, golden, borderline fail, dirty, shiny and color variants. Bring fixtures or describe placement variation honestly.
3. Map the station envelope, working distance, robot interference, cable routes, eye-safety zones for lasers, and ambient light sources (windows, welding arcs nearby).
4. Decide architecture, onboard smart 3D tools versus PC cloud processing; revisit smart vs PC.
5. Shortlist technologies, profiler vs stereo vs structured light vs industrial RGB-D against the eight buying questions above.
6. Run a feasibility trial, see the feasibility checklist; insist on your parts, not brochure parts.
7. Plan calibration and maintenance, who recalibrates after a bump? What artifact stays on site?
8. Freeze acceptance criteria, Z and XY metrics, false-reject budget, and cycle time including compute, then open the camera catalogue.

Many project teams also underestimate data hygiene. Point clouds and depth maps are large. Decide early whether you store raw clouds, downsampled meshes, or only numeric results. Storage and network design belong in the same conversation as sensor selection.

Finally, keep 2D in the toolkit. A parallel area scan view for print or code reading is often cheaper than forcing every decision into the 3D channel. Hybrid stations are normal, not a sign that you chose wrong.

Short glossary

Point cloud

Set of XYZ points representing the measured surface.

Depth map

Image-like grid where each pixel stores distance or height.

Profiler / laser triangulation

Height from a laser line viewed at an angle; often scanned with motion.

Stereo

Depth from disparity between multiple views.

RGB-D

Registered color (or intensity) plus depth from one device family.

Working volume

3D region where the sensor meets its stated performance.

Hand–eye calibration

Aligning camera coordinates with robot coordinates.

FAQ: 3D cameras

How this guide fits your selection

Use this guide when you are still deciding whether depth measurement is justified, before you compare 3D model options. Focus on surfaces, working volume, Z tolerance and robot or PLC integration.

Related reading lives under Knowledge.

Next steps

You now have the category model: 3D measures depth for shape, height, volume and pose, using stereo, profilers, structured light or RGB-D, only when geometry is the decision.

Related: Knowledge hub · Lighting · Smart vs PC.