CAD template vs camera-captured template for 3D matching accuracy

Every 3D matcher needs a reference shape, and there are two ways to get one: generate the template from the CAD file, or capture it from a real part with the same camera that runs the line. Teams argue about which is more accurate, and the argument ends with a number, not a preference. A CAD template carries no sensor noise, no lighting artifacts and no calibration error, which is why it wins most rows of the error budget — but it describes the part as designed, not the part as made. This comparison gives you the decision rule, the budget behind it, a worked example with the arithmetic shown, and the checks that settle the question for your parts before you commit.

The decision

Use the CAD-generated template as the default, and switch to a captured template only when the real part deviates from CAD by more than the captured template's own error:

Here d is the worst form deviation between real parts and the CAD surface, σt the noise left in the captured template, c the calibration residual of the capture setup, and p the lateral point spacing of the scene cloud. For a mid-range laser profiler at p = 0.5 mm with 0.06 mm repeatability on a fixed mount, the switch point is 0.078 mm. Ground and machined parts sit far below it; molded, cast, formed and natural parts sit above it, and there the captured template wins by three to four times, as the worked example shows. The field screen, one tenth of the point spacing, is deliberately conservative: it ignores the captured path's calibration residual, so it switches you slightly early, never late.

Why the CAD template usually wins

Four error sources tax a captured template and never touch a CAD template:

• Sensor noise. Every captured point carries the sensor's random error, and the matcher then measures scene noise against template noise. The CAD surface is exact geometry; sampled at the scene's spacing, it adds nothing random.
• Lighting artifacts. A capture embeds its day: highlights punch holes, dark surfaces drop points, and the matcher reads those holes as geometry. Mask low-confidence points before saving the template — the confidence image bitflags guide shows how to decode the flags and drop only the bad pixels.
• Calibration residual. A captured template inherits the extrinsics of the setup that took it. A fixed mount holds roughly 0.05 mm; a robot-held sensor carries about 0.1 mm after hand-eye calibration. The CAD file needs no frame to exist. When the depth map must reach the matcher in engineering units, the depth map to real distance calibration workflow is the part to get right.
• Coverage gaps. One capture is one viewpoint: occlusion shadows and steep-angle dropouts stay in the template. Filling them needs several captures merged — copy the gap-free point cloud merge pattern.

A fifth advantage is quieter: reproducibility. The same CAD file regenerates the identical template on every line, forever. A captured template must be re-taken after any sensor swap, recalibration or mount change, and two lines then hold two slightly different references until someone notices.

Where the captured template wins

The CAD file is nominal; parts are as-built. Once d passes the switch point, the matcher pulls each pose toward geometry the real part does not have, and that systematic error survives every average. The deviations that cross the line in practice:

• Sheet-metal springback after forming — tenths of a millimetre on a 100 mm flange.
• Injection-molded warp and shrink, worst in glass-filled resins.
• Castings, where draft, flash and parting-line mismatch are idealized away in CAD.
• Worn or reworked tooling, so running parts match no released revision.
• Printed parts matched as-built, and natural products — wood, food, leather — that never had a nominal form.
• No CAD at all: legacy and supplier parts. Capture several known-good parts and compare the templates; when the template-to-template spread passes 0.1 × p, no single template represents the family, so match on stable sub-features instead.

None of this makes capture inherently more accurate. It makes the CAD template systematically wrong, and a systematic error larger than the noise beats the noise every time.

The error budget behind the rule

Both paths end in the same matcher against the same scene cloud, so scene noise σ appears in both and cancels in the comparison. What differs is the template-side term:

As a class, laser profilers deliver Z repeatability of 0.02–0.1 mm at mid working distances and lateral spacing of 0.2–1 mm; projected structured-light cameras trade accuracy, typically 0.1–0.5 mm, for a single-shot full-field capture. Put your sensor's published repeatability in for σ and your configured spacing in for p — the point cloud x y z resolution settings guide shows where that setting lives and how to change it without distorting the cloud.

Worked example: two parts, one profiler

Example inputs: fixed-mount laser profiler, lateral point spacing p = 0.5 mm, repeatability σ = 0.06 mm, calibration residual c = 0.05 mm, single-frame template so σt = σ.

Template-side error, captured:

√(0.06² + 0.05²) = √(0.0036 + 0.0025) = √0.0061 = 0.078 mm

Part A, ground steel locator block, d = 0.02 mm.

• CAD: √(0.02² + 0.06²) = √0.0040 = 0.063 mm
• Captured: √(0.078² + 0.06²) = √0.0097 = 0.098 mm
• Decision: CAD, about a third tighter.

Part B, glass-filled nylon bracket, d = 0.30 mm.

• CAD: √(0.30² + 0.06²) = √0.0936 = 0.31 mm
• Captured: 0.098 mm
• Decision: captured, more than three times tighter.

Improve Part B's template by averaging 16 captures: σt = 0.06/√16 = 0.015 mm, the template-side term falls to √(0.015² + 0.05²) = 0.052 mm, and the matching error to √(0.052² + 0.06²) = 0.079 mm — four times tighter than the CAD path on the same part.

CAD versus captured at a glance

How to capture a template that beats CAD

1. Mount the sensor on a rigid stand, not on the robot. The fixed mount keeps c near 0.05 mm; hand-eye calibration is what doubles it to 0.1 mm.
2. Fixture the golden part in the orientation the matcher will see, and mask the fixture out of the cloud before saving.
3. Mask low-confidence points by decoding the flags, not by thresholding intensity.
4. Average 16 captures: 0.06 mm of repeatability becomes 0.015 mm.
5. Capture from a second angle to fill occlusion shadows, then merge the clouds.
6. Match the template's spacing to the scene's. A scene decimated to 0.5 mm gains nothing from a 0.05 mm template; when the scene resolution looks finer than you configured, check decimation first.
7. Validate: register 10 fresh parts and record the pose scatter; accept when the scatter stays under a third of your placement tolerance.

Choosing the sensor for either path

The template debate changes only the reference shape; the scene cloud sets the floor, because σ, p and c come from the sensor and the mount. Buy those first:

The MIS-DP3850-01P is a laser profiler for inline scanning as parts pass. The MV-DPS200P-02 is a projected structured-light camera, the natural bench tool for golden-part template capture with a single full-field shot. The MV-DLS300P-04 steers its laser with a galvanometer, which suits larger workcells. If operators need a synchronized 2D view for review, pair the station with an area scan camera and one of the C-mount machine vision lenses; that channel helps people, not this error budget.

For wide inline coverage, the MV-HLVS-16K100-L60 is a line-scan 2.5D system with 16K resolution, a 120 kHz sampling rate and a 600 mm emitting area.

Checks before you commit

• Measure d on five real parts: register each cloud to the CAD template and take the 95th-percentile signed distance of the worst part. Nominal tolerances understate as-built warp.
• Compare that d with √(σt² + c²) for your mount — about 0.05 mm fixed, about 0.1 mm on the robot wrist.
• Confirm the CAD revision matches running production, not the newest release.
• Strip features smaller than p from the template; the cloud cannot confirm them and they only add mismatch.
• Re-run the 10-part pose-scatter test after any sensor swap, recalibration or mount change.
• When match scores drop after a part change, check clouds and poses before blaming the template — the 3D fine matching input mismatch guide orders those checks.

When the usual answer is wrong

• "CAD is always more accurate" dies on Part B: 0.31 mm against 0.098 mm on the same sensor.
• "Captured is more real, so capture" dies on Part A: 0.035 mm of avoidable error plus a re-capture burden at every recalibration.
• The template is only half the pipeline. Deproject the depth with the correct camera's model — the aligned depth point cloud intrinsics guide picks the right one — because the wrong intrinsics shift the whole cloud by more than the entire template-side budget.
• On parts that vary unit to unit, neither pure source is ideal: register coarsely against CAD, then refine on stable faces of the scene cloud itself and exclude the regions that warp.

Where to go next

Work the budget for your own line: one published repeatability, one configured spacing, one mount constant, and the switch point falls out in a minute. Browse the 3D sensors category for the capture hardware, start at the camera step of the system configurator to build the capture station, or read the wider 3D vision topic for capture and registration workflows. For one specific part, send the CAD file and a sample cloud through the contact page and the numbers above get run against your parts.