Depth camera pixel size: estimate it from field of view and resolution

The depth data sheet lists a field of view and a depth-map resolution, and stays silent on the IR receiver's pixel pitch. You do not need it. The number that sizes features, sets gripper tolerances and predicts whether a defect spans enough pixels is the scene width one depth pixel covers at your working distance — the object-space pixel size. This guide derives it from the two numbers you already have, quantifies how it moves with distance, resolution and field of view, and works the arithmetic for a wide-angle depth camera at a 1000 mm working distance. It applies to ToF, stereo and structured-light depth cameras; line-scan 2.5D systems get their own variant further down.

The pitch is not the number you are after

Pixel pitch describes the receiver, in micrometres. Object-space pixel size describes the scene, in millimetres per depth pixel at a stated distance. A depth camera's output grid rarely matches its receiver one to one: stereo pipelines match image pairs and resample the result, ToF cameras demodulate phase per output pixel, and structured-light systems decode projected patterns into a disparity or height map. The depth output can be produced at a resolution different from the raw sensor, so even a published pitch would not tell you the scene sampling.

For feature sizing, resolution budgets, gripper offsets and mounting height, work with the object-space figure. If you need the physical pitch itself — sensor-format arithmetic, optical spot size — the depth camera pixel pitch guide covers how to recover it when a data sheet omits it. Everything below builds the scene figure from the two numbers a depth data sheet leads with: field of view and depth-map resolution.

The formula: field of view and depth-map resolution are enough

A depth stream follows pinhole projection, so three symbols do all the work:

• θ_h — horizontal depth field of view in degrees, quoted for the depth stream itself
• N_h — depth-map width in columns, at native output resolution
• d — working distance along the optical axis in millimetres

s(d) is the scene width one depth pixel spans at distance d, averaged across the row. The vertical figure uses the same form with the vertical FOV and the row count. A sheet that quotes one diagonal angle needs converting to horizontal first — the FAQ below works a 4:3 example.

Two finer points hide inside the average. The centre pixel always spans less than the average, by the factor tan(θ_h/2)/(θ_h/2 in radians) — 1.26 at the half-angle in the worked example — and an edge pixel spans the centre value multiplied by sec²θ, the tangent stretch that pulls wide-angle corners apart. The projection is the same geometry a 2D lens follows:

For a conventional 2D lens the same scene quantity comes from magnification instead of FOV; the object space pixel size guide derives it that way.

Worked example: 87° FOV, 1024 columns, 1 m working distance

Take a wide-angle depth camera: θ_h = 87°, native depth map N_h = 1024 columns, mounted at d = 1 m = 1000 mm.

1. Half-angle: 87° / 2 = 43.5°, and tan(43.5°) = 0.949.
2. Scene width: W = 2 × 1000 mm × 0.949 = 1898 mm.
3. Average pixel size: s = 1898 mm / 1024 px = 1.85 mm/px.
4. Angular step: Δθ = 87° / 1024 = 0.085° = 0.00148 rad. The centre pixel spans 1000 mm × 0.00148 = 1.48 mm; an edge pixel spans 1.48 mm × 1.90 (sec² at 43.5°) = 2.82 mm.

Plan around 1.85 mm per pixel on average, 1.5 mm at the centre and 2.8 mm at the frame edge. Distance scales every value linearly:

Inverted for design, the formula sets a resolution budget: to hold 1.5 mm/px or better at that reference distance with 87° optics you need N_h ≥ 1898 mm / 1.5 mm = 1266 columns, so a 1280-column depth map just clears the target. When you start from a part and a defect size instead of from a data sheet, the resolution calculator below runs that inverse problem.

What moves the number, and by how much

• Distance, linear. Doubling the working distance doubles the step: 1.85 → 3.71 mm/px from 1 m to 2 m in the example.
• Column count, inverse. Halving the native columns doubles the scene step at the same distance.
• Field of view, through tan(θ/2). Swapping a 70° optic for a 90° optic at equal distance and column count widens each pixel by tan(45°) / tan(35°) = 1.43 — 43% coarser sampling.
• Input errors. The costly mistakes are inputs, not optics. An RGB field of view used for the depth stream overstates coverage, because on stereo pairs the depth image covers less than the colour picture; the stereo depth field of view article quantifies the gap. An upsampled output grid divided as though native improves the figure on paper only. And a depth value that is slant range rather than axis distance needs converting first — the depth value vs true distance guide shows the conversion.

The small-angle shortcut s ≈ d·θ_h(rad)/N_h is worth knowing and worth distrusting: it under-reads the average by 4% at a 20° half-angle, 10% at 30° and 25% at the example's 43.5°. Keep the tangent form whenever the FOV is wide.

Using the number: features, tolerances and accuracy budgets

The pixel rule converts sampling into decisions: about 3 pixels across a feature to detect it, 5 to measure its size. At 1.85 mm/px:

• a 10 mm feature spans 10 / 1.85 = 5.4 pixels, so it measures cleanly;
• the smallest detectable mark is about 3 × 1.85 = 5.6 mm;
• ±1 pixel of repeatability is ±1.9 mm, and a subpixel edge fit at ±0.3 pixels is about ±0.6 mm.

The camera resolution for defect detection guide works the pixel rule from the part side, and the edge position accuracy article shows when the subpixel fraction is realistic. Object extent read from a depth bounding box inherits the same sampling — the 2D bounding box to 3D guide gives that estimate's error behaviour.

For robot placement, divide tolerance by sampling. A ±4 mm pick tolerance at 1.85 mm/px is a ±2.2 pixel budget, which a calibrated system meets; ±1 mm is ±0.54 pixels, which asks for subpixel methods or a shorter working distance rather than luck. Turning the calibrated pixel grid into robot-base coordinates is its own pipeline — the camera calibration walkthrough from pixels to a robot base transform follows it end to end.

Where to trust the estimate and where to calibrate

For selection and layout, the formula carries a few percent, because FOV and column count are the data sheet's own figures. Three effects move a real rig away from paper. Rectilinear wide-angle optics bend corners, so edge pixels stretch a few percent beyond the tangent model. Production depth maps are cropped, which changes coverage but not the scale of each pixel, or resampled, which changes the grid without adding information — the stereo depth output resolution article shows why size settings seem ignored. And quoted FOVs differ in convention: diagonal versus horizontal, or defined at a particular distance.

When pixels feed acceptance decisions, calibrate the unit itself: place a target of known size at the working distance and fit mm per pixel from it. At 1.85 mm/px a 400 mm target spans 400 / 1.85 = 216 columns, enough for a proper checkerboard at the reference distance — which is why target size, not the camera, rules calibration range at longer distances, as the checkerboard target size article works out.

Line-scan 2.5D and laser profilers: one number per axis

Two depth geometries refuse the single division. A line-scan 2.5D system samples along the sensor line with optics and across travel with motion. The MV-HLVS-16K100-L60 pairs 16K profile resolution with a 120 kHz sampling rate and a 600 mm emitting area; its travel-axis step is conveyor travel per profile, a number no pixel pitch would give you.

A laser profiler such as the MV-DP4180-03P takes its profile-axis scale from the triangulation optics over the measurement span and its travel axis from encoder motion. Derive the two axes separately and keep them separate in the tolerance budget.

Choose the sensor from the sampling budget

1. Fix the sampling target from the task: feature size divided by 3 for detection (5 mm marks → 1.7 mm/px), tolerance divided by 0.3 for measurement (±0.6 mm → 2.0 mm/px).
2. Fix the working distance from the mounting envelope; the 3D camera working distance and positioning guide covers how close is practical.
3. Compute the column budget: N_h ≥ 2·d·tan(θ_h/2) / s_target — at the 1000 mm reference and 87°, 1.5 mm/px needs 1266 columns.
4. Check coverage the other way: W(d) = 2·d·tan(θ_h/2) must wrap the part plus handling margin — 1898 mm at the reference distance in the example.
5. Shortlist units whose depth FOV and native depth columns clear both numbers, then measure once on the rig.

The same arithmetic sizes any area depth camera, because ToF sensing cameras, projected structured-light cameras and laser galvanometer stereo cameras all quote a depth field of view and a native depth-map size.

Browse the wider parts bin in the 3D sensors category or filter by resolution and series in the camera finder.

Five checks before you commit

• Recompute from the depth stream's own FOV and native column count for the exact model, never from a colour camera's table row.
• Scale with distance deliberately: at 2 m every worked figure doubles, at 0.5 m it halves.
• Keep measurement zones near the frame centre, where a pixel spans the least scene; corners run 1.9× coarser at a 43.5° half-angle.
• Measure once with a known-length target and expect single-digit percent agreement with the paper value; a bigger gap means a wrong FOV source or column count.

When the budget is settled, take it to the configurator: start at the camera step and apply the FOV and resolution figures you derived. Related depth-camera geometry — working distance, ROI, resolution settings — is gathered on the 3D vision topic page.