3D camera specs explained: precision, trueness and spatial resolution

A 3D camera datasheet quotes point precision, planarity error, trueness and spatial resolution. Most buying mistakes come from reading all four as one number called accuracy. This guide is for engineers specifying structured-light, laser-profiling, stereo and time-of-flight sensors for picking, gauging and surface inspection. You will learn what each spec measures, how to compute the spatial resolution at your own working distance, how to build an error budget from precision and trueness, and how to compare two datasheets that use different words.

What do precision and trueness mean on a 3D camera datasheet?

Point precision is the random capture-to-capture spread of a single measured point, quoted as a 1σ distance. Trueness is the systematic offset between the measured geometry and the true part, quoted in millimetres or as a percentage of dimension. Accuracy requires both to be small.

Hold a part still and capture it twenty times. The same surface point lands in a slightly different XYZ position on every capture because of random noise, and the 1σ radius of that cluster is the point precision. The average of the twenty captures can still sit a fixed distance from where the surface truly is: that offset is trueness, and it repeats identically in every capture.

The two failure modes look different on screen. Precise but untrue gives a clean point cloud with a wrong pose or a wrong dimension, the dangerous case because the data looks trustworthy. True but imprecise gives noisy captures whose average still lands correctly, so averaging or filtering recovers it. Only both together give accuracy.

Averaging helps exactly one of the two. Four independent captures cut the random spread to about half, sixteen cut it to a quarter, but a systematic 0.4 mm scaling error survives ten captures and a hundred unchanged. That asymmetry is why trueness usually decides gauging applications while precision decides picking.

Both figures are quoted for a defined calibrated measurement volume, and trueness changes with position inside it. For stereo systems the volume effectively decides the accuracy available, as industrial 3D camera selection: stereo, ToF or structured light works through.

What is spatial resolution, and what do you get at your working distance?

Spatial resolution is the physical surface area one point samples. It is roughly the calibrated field of view divided by the points across it, so a 320 mm field sampled by 2048 points gives 320 / 2048 = 0.16 mm per point at the calibrated working distance.

Sensor resolution and spatial resolution are different specs. Sensor resolution is the pixel count of the image sensor: every pixel that contributes a 3D point becomes one XYZ sample. Spatial resolution is the width of surface one sample covers, and it is set by the calibrated field of view, the point count and the distance, which is why datasheets always tie it to a working distance.

Worked example, using inputs typical of a mid-range structured-light camera: 2048 points across a 320 mm calibrated field at a 500 mm working distance.

• s = 320 mm / 2048 = 0.156 mm per point, about 0.16 mm.
• A 12 mm notch spans 12 / 0.156 ≈ 77 points, far above the 8-10 points you want across any feature whose size must be trusted.
• The same capture with 2×2 subsampling: s = 320 mm / 1024 = 0.31 mm, and the notch spans 12 / 0.31 ≈ 38 points. Still measurable; a 3 mm chamfer now spans under 10 points and starts to slip away.

For a fixed lens the field of view widens with distance and each point's footprint widens in proportion: at double the distance, 0.16 mm per point becomes roughly 0.32 mm. Use that scaling to size the measurement volume you need. It is not a licence to move the camera outside its calibrated volume, where the printed figures stop being valid.

Which sensor pixels carry real 3D data, and how projector pixels map onto them, is the subject of structured light 3D cameras: projector pixel mapping to 3D points.

Why is there no Z resolution on the datasheet?

A single depth number describes one axis, but a 3D point is measured in X, Y and Z at once and its error is a three-dimensional distance. Datasheets that quote one noise figure therefore use point precision, the 1σ Euclidean spread of a single point across captures.

The term depth resolution is used inconsistently: some vendors mean the axial step of the depth output, some mean Z-axis noise, and either way a Z-only figure hides lateral error entirely. Point precision collapses all three axes into one distance you can budget with. When a datasheet offers only a Z accuracy figure, ask for the pair — a precision figure and a trueness figure — before comparing it against anything else.

What does dimension trueness error cost you on a real part?

Dimension trueness error is quoted as a percentage of the measured size. A sensor with 0.2% trueness reads a 200 mm dimension as 199.6 to 200.4 mm, a 100 mm dimension as 99.8 to 100.2 mm and a 300 mm dimension as 299.4 to 300.6 mm.

The error in millimetres grows with the largest dimension you trust from the cloud: 0.002 × 400 mm = 0.8 mm. A camera that handles a 30 mm boss well can still misplace a 400 mm tote by most of a millimetre. Check the largest dimension in the scene, not the smallest. Global planarity trueness is the shape version of the same bias: a physically flat surface bows in the point cloud across the field of view, and a bowed tote base tilts every pose computed from it.

Now build the budget for a placement task with these inputs:

• Tolerance: the fixture accepts the part with ±1.0 mm of total play.
• Point precision: 0.10 mm 1σ, so the 3σ random contribution is 3 × 0.10 = 0.30 mm.
• Dimension trueness: 0.2% on the 200 mm span that locates the part, so 0.002 × 200 mm = 0.40 mm systematic.
• Combined root-sum-square: √(0.30² + 0.40²) = √0.25 = 0.50 mm; if both errors land in the same direction the worst case is 0.70 mm.

Decision: against a ±1.0 mm tolerance the root-sum-square error is exactly half of it, leaving the other half for robot repeatability, fixturing and part variation — a workable budget. Tighten the tolerance to ±0.5 mm and the same sensor fails on trueness alone. The fix is in-place trueness correction or a tighter sensor, not more pixels.

The same split diagnoses installed cells: a pick that lands offset the same way every cycle is a systematic, trueness-style error, while a pick that scatters is precision. The workflow in diagnose an eye in hand pick point offset separates the two on a live cell.

Which spec decides your task: picking, gauging or surface inspection?

Match one spec to each task. Random part picking lives on point precision and spatial resolution, flat-surface and pallet detection on global planarity trueness, dimensional gauging on dimension trueness, and surface-quality work on local planarity precision.

Point precision needs repeat captures, and production rarely gives you more than one shot. Local planarity precision is the one-capture equivalent: the roughness of a small patch of points against its own fitted plane, quoted for a patch of a stated size.

As a worked example, take a 0.05 mm local planarity figure on a 50 × 50 point patch. At 0.16 mm spacing the patch covers 50 × 0.16 = 8 mm of surface, and a physically flat face shows ±0.15 mm of ripple at 3σ (3 × 0.05 mm). A 0.2 mm-deep defect sits at the noise edge and a 0.1 mm defect is invisible, so the shallowest defect you must report should be at least three times the local planarity figure.

The same logic bounds global flatness. On a 400 mm tote base, 2 mm of bow inside a 100 mm pick height margin is 2% of it and harmless; inside a 5 mm margin it consumes 2 / 5 = 40% and will cost picks. Read the planarity spec against the margin the cell actually has.

How do the specs differ between structured light, laser profiling, stereo and ToF?

Structured-light area cameras publish the full precision-and-trueness vocabulary, laser profilers quote per-line precision and accuracy over a profile range, stereo depth error grows with distance squared, and ToF trades detail for range. The same part asks different specs from each family.

Structured light returns a dense full field in one capture, so point precision, both planarity figures and both trueness figures apply to it directly. The MV-DPS200P-02 is a projected structured-light 3D camera of exactly this kind.

Laser profiling reports precision along each scanned line and accuracy over the calibrated profile range, while the along-conveyor axis is set by the encoder travel between lines. The MV-DP4180-03P is a 3D laser profile sensor of the DP4000 series for this style of scanning.

Stereo triangulates depth from two views, and its depth error grows with the square of the distance ratio: a pair that resolves 0.5 mm at 500 mm resolves about 2.0 mm at 1000 mm, because (1000 / 500)² = 4. The MV-DLS300P-04 is a laser galvanometer stereo camera: a steered projector paints texture onto featureless parts so the stereo pair has something to match.

ToF leads with range: metres of working volume with millimetre-to-centimetre precision, so its datasheets emphasise range and field of view rather than sub-millimetre planarity. The MV-DT01WMU is a ToF sensing camera for range-first jobs such as volume and presence detection.

For flat parts on a moving line, line-scan 2.5D systems build the height map from motion: the MV-HLVS-16K100-L60 covers a 600 mm emitting length at a 120 kHz sampling rate with 16K points across the profile. For a stationary shiny part, the MV-PSA-12UM-C4D30-WDM pairs a 12 MP camera with a 300 mm dome light for diffuse full-field height inspection.

When the tolerance is tighter than about 0.1 mm, structured light is normally the family that reaches it, and measuring a foot from a photo: why 0.1 mm needs structured light shows why weaker triangulation falls short.

Why is a higher pixel count not a better 3D camera by itself?

Pixel count sets how many points you sample, not how truthfully or steadily each point is measured. A 2 MP sensor with better precision and trueness outperforms a 12 MP sensor whose points are noisy or systematically warped, especially once subsampling or interpolation is in play.

Three effects decouple pixel count from measurement quality. First, subsampling: 2×2 modes quarter the point count and 4×4 modes cut it sixteenfold, trading density for speed and file size, so a full-resolution headline may not describe what the cycle actually runs. Second, interpolation: an upsampled depth frame holds more pixels but no more measurements, and its real accuracy is settled by the underlying data, which upsampled depth frames: alignment resolution vs accuracy quantifies. Third, the same trade exists in 2D imaging, where optics and per-pixel quality rather than the megapixel headline set visible detail, as color camera spatial resolution: beyond the pixel count explains.

Decide on the two numbers that do the work: points across your smallest feature (spatial resolution), and the spread and bias of each point (precision and trueness). Pixel count only sets the ceiling those numbers can reach.

How do you compare two 3D camera datasheets fairly?

Two datasheets are comparable only when the measurement conditions match: the same working distance, the same surface reflectivity, the same patch size in millimetres for planarity, the same statistic and confidence for trueness, and the same exposure mode. Normalise those five before ranking any numbers.

• Working distance. Compare spatial resolution at your part distance, not at the vendor's demonstration distance, and recompute with s = FOV / N whenever the field differs.
• Surface. Printed precision assumes a matte, diffusely reflecting target; a dark or polished part lowers the signal and widens the spread, so ask for a capture on a part like yours.
• Patch size. A 50 × 50 point patch on a fine-pixel sensor covers fewer square millimetres than on a coarse one, so convert patches to square millimetres before comparing planarity.
• Trueness statistic. Some vendors print a mean, some a percentile, some a bound over the whole volume after thermal soak; convert everything to a worst case over your volume.
• Exposure mode. Single-shot precision and multi-exposure precision differ; quote the mode the cycle will actually run.

With conditions matched, build the budget from the worked example above and pick the camera whose combined error sits under half your tolerance. Confirm the calibrated volume covers every part pose, because trueness is quoted for that volume. Keep the asymmetry in mind: the systematic trueness of many structured-light cameras can be re-characterized in place with a calibrated artifact after mounting, while precision stays as measured.

Shortlist from the 3D sensors category or the camera finder, then take the error budget to the configurator:

The decisions, one line each: picking lives on point precision and points across the feature, gauging lives on percentage trueness times the largest dimension, surface work lives on local planarity against the shallowest defect, and pixel count only sets the sampling ceiling. Budget both error components to half the tolerance and the vocabulary stops being a trap. The rest of a 3D build, from sensor choice to calibration and mesh export, is collected on the 3D vision topic page.