Depth camera pixel pitch: get it without a datasheet

Depth camera datasheets print resolution and optical format, yet omit pixel pitch more often than any other sensor number. You rarely need to wait for one: pitch falls out of a single division — active sensor width divided by the pixel count across that width. This guide gives the arithmetic, typical active areas for every common optical format, a worked example with a cross-check, the binning correction, and the three depth-camera calculations where the number changes your answer: per-pixel footprint, field of view and stereo depth noise.

Pitch, resolution and active area

Pixel pitch is the center-to-center spacing between neighboring photosites on the sensor, quoted in micrometers. Resolution is the pixel count, written N × M. Active area links the two: a 1280 × 960 array at 6.9 µm pitch occupies 8.8 × 6.6 mm. Each quantity predicts the other two, which is why a missing one can be recovered.

On a depth camera, the pitch describes the sensor that samples the infrared image — the speckle field of a structured-light projector, the two images of a stereo pair, or a time-of-flight pulse. Depth is computed per sensor pixel, so pitch sets the lateral sampling of the point cloud: two features closer together than one pixel's footprint at the object cannot be separated in X and Y, whatever the Z precision. That is the same pixel rule used to size a 2D camera for defect detection, applied one dimension short.

Watch the vocabulary: pixel size gets used for both the sensor-side pitch and the object-side footprint of one pixel. The division below gives the sensor-side pitch; the footprint follows from it two sections down.

The division, and where the sensor width comes from

Something else is almost always printed when the pitch is not. Three sources for the sensor width, in order of preference:

1. The optical format. Listings say 1/2-inch, 1/3-inch, 2/3-inch. The inch label is a legacy of vidicon camera tubes, not a measurement: a 1/2-inch sensor has an 8 mm diagonal, not 12.7 mm. Take the width from the table below instead of converting the fraction.
2. SDK or camera metadata. Many depth camera SDKs report sensor and pixel information through their device query. Where your SDK prints a pixel size, it overrides the table.
3. The lens. A lens built for 2/3-inch covers an 11 mm image circle, and the active area sits inside it. That brackets the width; treat it as a bound, not as the number.

These are the typical active areas quoted across the industry for each format. Individual sensors land a few percent away, so a pitch derived here holds to roughly ±5 percent; the sensor's own datasheet beats the table when one exists.

Worked example: 1280 × 960 on a 2/3-inch sensor

Take a depth camera that streams 1280 × 960 and carries a 2/3-inch sensor. The table gives an 8.8 mm active width:

p = 8.8 mm / 1280 px = 0.006875 mm ≈ 6.9 µm

Cross-check on the other axis: 6.6 mm / 960 px = 0.006875 mm. Both axes agree, so the pixels are square and the format assumption holds. Carry 6.9 µm into every calculation below.

Now the case that catches people. A 1280 × 1024 array on a nominal 1/2-inch sensor divides to 6.4 mm / 1280 = 5.0 µm across, but 4.8 mm / 1024 = 4.7 µm down — about 7 percent apart. Square pixels cannot satisfy both, so an assumption is wrong: the real active area differs from the nominal value, or the array is not square-pitched. Use each axis's own pitch for that axis's math, and pull the exact active area if the difference feeds a measurement.

Correct for binning and streamed depth resolution

The division uses the native active array, but depth cameras rarely stream it raw. Many offer 2 × 2 binning, which turns 1280 × 960 into 640 × 480 and doubles the spacing of the streamed pixels. Some decimate instead, and the depth processing itself can output at a reduced size — the reason a resolution change sometimes looks ignored on a stereo camera; the stream-resolution article untangles that case.

• Binning merges photosites: effective pitch = native pitch × binning factor. At 6.9 µm native, 2 × 2 binning gives 13.8 µm per streamed pixel and doubles the footprint, to about 4.9 mm per pixel at a 1 m standoff with the 2.8 mm lens used below.
• Decimation skips pixels: pitch is unchanged, but coverage is lost between the kept pixels.
• Cropping changes neither: only the extent shrinks.

Use native pitch for optics math — depth of field, circle of confusion, diffraction, all covered in working f-number and magnification — and effective pitch whenever a streamed pixel converts to an object size. The same native-versus-streamed split governs point-cloud spacing, which the x-y-z resolution article covers alongside the distortion traps.

Area-scan depth cameras — time-of-flight and projected structured-light designs, such as the MV-DT01WMU and the MV-DPS200P-02 — all sample the scene on one sensor, so the same division and the same binning logic give the pitch behind their depth maps:

Where pitch changes your depth math

Fix an illustrative optics set — a 2.8 mm wide-angle lens, a 50 mm stereo baseline, the 6.9 µm pitch from the worked example — and three numbers follow. Substitute your own values; the arithmetic is identical.

Per-pixel footprint. For a rectilinear lens, the object-plane width of one pixel is:

At a 1.0 m standoff: ΔX = 1000 mm × 0.006875 mm / 2.8 mm = 6.875 mm² / 2.8 mm ≈ 2.46 mm. Each pixel covers about 2.5 mm across; at 0.5 m it covers 1.23 mm, at 2 m about 4.9 mm. The footprint is linear in both distance and pitch, so halving either halves it. This is also how you size a calibration target: a 100 mm checkerboard square spans 100 mm / 2.46 mm ≈ 41 pixels at 1 m, which is why targets that look generous on the bench go sparse at range — the checkerboard target article works the same problem from the target side.

Field of view. Multiply through by the pixel count:

The geometry — active width, lens, standoff, field — runs both directions: give the lens focal length calculator the field you need at the distance you have, and it returns the focal length.

Stereo depth uncertainty. For a stereo or speckle-triangulation camera, depth error per pixel of disparity noise grows with the square of distance:

One pixel of disparity error at 1 m: σ_z = (1000 mm)² × 0.006875 mm / (2.8 mm × 50 mm) = 6875 mm³ / 140 mm² ≈ 49 mm. Practical correlators locate the match peak to a fraction of a pixel — 0.1 to 0.5 pixels is typical — so expect roughly 5 to 25 mm of depth noise at 1 m with these optics, and about 196 mm per pixel at 2 m, four times worse because of the z². Doubling the baseline from 50 mm to 100 mm halves every figure: the baseline, not the pitch, is the strongest lever at range, as the long-range stereo article shows with measured numbers.

A sanity check ties it together. With f = 2.8 mm and B = 50 mm, a point at 1 m has a disparity of f × B / z = 0.14 pixels — far below one. These cameras work there only by correlating a projected pattern with sub-pixel interpolation, which is why structured light and speckle dominate indoor depth while raw feature matching does not.

Time of flight differs. A ToF pixel's depth comes from the light's round trip, so the σ_z formula does not apply to it; the pitch still sets the ΔX footprint exactly as above. Standoff then becomes plain arithmetic: to sample 5 mm laterally with the worked-example optics, z = 5 mm × 2.8 mm / 0.006875 mm ≈ 2.0 m. The 3D camera working distance guide collects the rest of that positioning decision.

When not to use a pitch estimate at all

Profilers quote what matters directly. A laser profiler's triangulation geometry is fixed at the factory, so its printed Z resolution, Z linearity, profile rate and scan width already contain the sensor pitch; deriving it would only add error to numbers you were handed. For a 3D laser profile sensor such as the MV-DP4180-03P, design from those printed Z figures.

Line-scan 2.5D systems print their scan geometry the same way. The MV-HLVS-16K100-L60 reads 16K across the scan line at a 120 kHz sampling rate, with a 600 mm emitting area length; per-pixel size across the scan is the emitting width divided by the active profile count — a division someone has already done for you.

Checks before you rely on the derived number

• Both axes divide to the same pitch within a percent or two. A spread above about 5 percent means the format or square-pixel assumption is off — the 7 percent case above.
• The streamed resolution equals the native active array. If binning or a reduced depth output is active, use effective pitch for anything object-side.
• The width came from the active area, not from the die size or a literal conversion of the inch fraction.
• Distortion: beyond roughly 100° of field angle, a wide lens bends the geometry and the constant-footprint map weakens toward the edges; a fisheye needs its projection model instead of one linear ratio.
• Metrology use: the table's ±5 percent may be wider than the job allows, so get the sensor's exact active area through the camera manufacturer's support channel.

The decision

For an area-scan depth camera with no printed pitch: divide the typical active width for its optical format by the native width pixels, cross-check on height, and switch to effective pitch when binning is on. That lands within about 5 percent — enough for footprint, field-of-view and standoff decisions. For stereo depth-noise estimates, the baseline and the z² growth dominate, so pitch precision is secondary. For profilers and line-scan 2.5D systems, skip the estimate and design from the printed Z figures.

From here: filter candidates in the camera finder, browse the 3D sensors category, or step through the system configurator when the camera is one part of a larger build. More depth-camera geometry lives on the 3D vision topic page, and the contact page reaches the applications team when a derived number is about to feed a production decision.