A depth measurement at 5-15 cm sounds simple until you read depth camera specifications. The modules built for rooms and robots start at 30 cm or more, and the ones that focus closer trade away the accuracy you need. This guide is for anyone building an inspection or measurement station that must separate 1 mm depth differences at a 50-150 mm standoff. It derives what compact stereo can actually resolve at this distance, compares the depth sensing classes, and shows why a laser line profiler clears the spec with room to spare — plus a camera-and-laser build if machine-vision parts are already on your shelf.
The decision, class by class
Commit first: a hard 1 mm pass/fail at 50-150 mm standoff is a triangulation job, not a depth camera job. Laser profilers clear it with margin to spare. Compact stereo sits exactly on the line. ToF and most structured-light cameras never focus that close.
| Technology | Typical minimum standoff | Depth precision near 100 mm | Verdict for 1 mm at 50-150 mm |
|---|---|---|---|
| Compact stereo depth camera | 70 mm, newest models lower | about 1% of distance, roughly a millimetre | Marginal: from 70 mm up, no margin |
| Direct ToF camera | 0.2-0.5 m | millimetre to centimetre class | No |
| Projected structured light | 0.2 m or more on most units | sub-millimetre to a few millimetres | Suited to the 100-150 mm half |
| Laser line profiler | tens to a few hundred mm | micrometre class | Yes, with 50-fold margin or better |
| Camera plus line laser | set by lens close focus | sub-millimetre, calibration-bound | Yes, if you can build and calibrate |
If you are weighing these classes for a whole cell rather than one measurement, the guide to stereo, ToF or structured light selection covers what each class is for.
Why compact stereo lands on the limit
Stereo depth error follows from triangle geometry:
σZ is depth uncertainty, Z the standoff, B the baseline between the two cameras, f the focal length in pixels and σd the disparity matching error. Two consequences dominate close in. A module that focuses at 70 mm needs very wide optics, and a wide field shrinks f in pixels: the same physical focal length covers far more scene, so each pixel of matching error turns into more depth error. Separately, calibration, synchronization and lens distortion put this class's systematic accuracy near 1% of distance — at the mid-window standoff that is about a millimetre, no matter how well the correlator matches pixels.
The hard stop is the minimum distance. Gripper-class short-range stereo is specified from 70 mm upward, so your 50-70 mm zone is out of range for current units — though the newest releases push the minimum under 3 cm on some models. Baseline choice drives the random error directly, as the guide to choosing a stereo baseline distance develops, and the same Z² behaviour governs long range stereo depth accuracy. Where the camera sits relative to the part — standoff, tilt, field placement — is a discipline of its own, covered in the 3D camera working distance and positioning guide.
Worked example: stereo depth error with a 20 mm baseline at the mid-window standoff
Inputs. Standoff Z = 100 mm, the middle of your 50-150 mm window (your project). Baseline B = 20 mm, typical for compact close-range modules built to fit a gripper. Depth image 1280 px wide across a 90° horizontal field — assume a wide field, because close-focus stereo modules ship wide-angle optics — giving f = 1280 / (2 × tan 45°) = 640 px. Disparity error σd = 1 px for integer matching, or 0.2 px with subpixel refinement on a textured part.
Rule. Depth uncertainty grows with distance squared and shrinks with baseline and focal length: σZ = Z² × σd / (B × f).
Arithmetic. At 1 px: σZ = (100 mm)² × 1 px / (20 mm × 640 px) = 10 000 mm²·px / 12 800 mm·px = 0.78 mm. At 0.2 px: 2 000 mm²·px / 12 800 mm·px = 0.16 mm. Depth pipelines often emit half-width maps — the stereo depth output resolution guide explains why size settings seem ignored — and at 640 px wide, f drops to 320 px, doubling the 1 px figure to 1.56 mm.
Result and decision. Random precision at this standoff lands at 0.16-0.78 mm, numerically under your 1 mm, but with no headroom once the systematic term near 1% of distance (about a millimetre here) joins it. Treat compact stereo this close in as guidance-grade: right for finding and posing parts, not for a hard pass/fail at that tolerance. The working threshold: a spec of 3 mm or looser, stereo clears; any tighter, triangulate.
The answer that clears the spec: laser line profilers
A laser profiler projects a line across the part, images that line from an angle and triangulates a height for every point along it. Sweep the part through the light plane — conveyor, motorized stage or robot — and each frame contributes one dense X-Z profile to the cloud. Nothing in the method needs surface texture, so matte plastic and machined faces measure equally well. At 50-150 mm standoffs, sensors of this class quote line repeatability from a few micrometres to a few tens of micrometres; even a loose 20 µm clears your step budget fifty-fold. Point spacing along the line falls in the 10-100 µm band, so a millimetre feature spans dozens of samples.
Three numbers on the series pages decide the fit: standoff (nominal sensor-to-line distance), Z-range (the measurable height band around it) and profile width. Pick a profiler whose specified standoff brackets your real one; a sensor built for 300 mm standoff will not measure cleanly at the shorter end of your window.
The shop's laser profile sensors come in three series — DP2000, DP3000 and DP4000 — plus binocular single-line variants and line-scan 2.5D systems for conveyor work. For a complete inspection built on one of these profilers, the surface defect detection with a 3D laser line profiler guide walks the whole pipeline.
- Type
- Laser profiler
- Type
- Laser profiler
Full-frame alternative: projected structured light
A profiler needs motion to cover an area. When the part must be captured in a single exposure — no axis available in the cell — projected structured light is the class: a coded pattern goes over the whole field and the camera returns depth per pixel in one shot. Class depth noise at short standoff runs from sub-millimetre to a few millimetres, and most units specify minimum ranges of 0.2 m or more. That makes this the option for the 100-150 mm half of your window, where one-shot area coverage beats building a motion axis. Why the finest surface metrology keeps landing on this method, and what 0.1 mm actually demands of structured light, is worth a read before you commit to a number tighter than yours.
A projected structured-light 3D camera such as the MV-DPS200P-02 is the class representative: full-frame depth from coded pattern projection.
- Type
- Stereo / structured light
Build it from parts: camera plus line laser
If machine-vision parts are already on the shelf, a triangulation rig is a weekend of work. Mount an area scan camera at the mid-window working distance, project a line laser at 30-45° to the optical axis, and calibrate image pixels to height. A height step displaces the imaged line sideways; even at a shallow 30° angle a millimetre step moves the line tens of pixels, so your tolerance resolves with room. The capture-and-calibrate procedure is laid out in the guide to creating a 2D image from laser triangulation, and the laser line extraction guide keeps the measured line clean.
Choose the lens from the field geometry:
A 1/1.8-inch sensor, 7.2 mm wide, covering a 30 mm strip at 100 mm WD gives m = 7.2 / 30 = 0.24 and f = 100 × 0.24 / 1.24 = 19.4 mm — pick a 20 mm lens from the machine vision lenses category. At 3072 px across that 30 mm strip, the pixel scale is 30 / 3072 ≈ 9.8 µm, so the line shift from a millimetre step spans dozens of pixels. The camera finder filters by interface, resolution and shutter, and the camera configurator steps through the rest.
Close focus is the usual snag: this close in, many lenses sit near the end of their focus travel. The fixes are a longer focal length, a close-focus variant or a thin spacer ring — the same logic as fixing a wide camera that will not sharpen at close range.
Set the geometry on paper before drilling anything: the focal length calculator solves the same field equation for your sensor and standoff.
Checks before you order
Run these before hardware money moves:
- Step artifact. Measure a milled 1 mm and 2 mm step, or gauge blocks, at 50, 100 and 150 mm standoff, 20 repeats each, and report the spread — not the best run.
- Sampling. Keep profile pitch well inside the smallest height feature; the laser triangulation scan spacing guide shows how sparse lines stretch a cloud into lies.
- Your surfaces. Fuzzy edges and thin walls drop out at this range, as the close range depth dropout guide explains; shiny metal needs the laser tilted away from the specular return.
- Motion budget. On a scanned profiler, axis jitter and trigger latency belong inside the tolerance budget; keep stage repeatability an order of magnitude under it.
- Mechanics. Micrometre-class repeatability assumes a rigid bracket and settled temperature; aluminium drifts a few micrometres per degree.
- Data path. Profilers stream profiles continuously and take hardware triggers; confirm your controller or PC keeps up at the profile rate you plan.
When your real requirement is looser
If your 1 mm figure actually means detecting that a step exists, a single profile line, or even an oblique 2D shot, answers it — no scan axis needed. If the deliverable is robot guidance rather than metrology, a compact stereo module at 100-300 mm sits comfortably inside gripper tolerance; the 2D or 3D camera for conveyor pick and place guide sets that decision out. If the standoff can move out to 0.3-1 m while the tolerance stretches to a few millimetres, full-frame structured light or ToF become the low-effort answers. And when depth-image edges vanish against a part boundary, the guide on why depth image edges vanish at long standoff explains what the depth pipeline is doing at that boundary.
Start in the 3D sensors category with your standoff, strip width and tolerance in hand, or send the three numbers through the quote request and we will size the profiler series and trigger chain for you. The 3D vision topic page collects the rest of the depth guides.
Questions
Can I just mount a depth camera further back from the part?
You can, but random error grows with the square of distance. Moving from 100 mm to 150 mm standoff multiplies it by (150/100)² = 2.25, turning a 0.78 mm figure into about 1.8 mm. Backing off only helps when the tolerance loosens faster than the error grows; for a hard step-height spec that tight, it moves the wrong way. Mount as close as the module's specified minimum allows.
What repeatability do laser profilers reach at a 100 mm standoff?
In this standoff class, industrial laser profile sensors quote line repeatability from about 1 µm at the shortest standoffs to tens of micrometres, with point spacing along the line in the 10-100 µm band. Even a loose 50 µm figure clears a millimetre spec by a factor of 20. Compare series on standoff, Z-range and repeatability together rather than any one alone.
Do I need a global shutter camera for laser triangulation?
Only if the part moves while the line is exposed. A rolling shutter reads the sensor row by row, so a line captured during travel comes out skewed. At a fixed station with the part still, rolling shutter images the line correctly and costs less. If parts scan past the camera, use a global shutter camera or a profiler triggered by encoder.
How do I prove 1 mm performance before installation?
Measure a reference artifact with known steps — 1 mm and 2 mm gauge steps or a calibrated standard — at each planned standoff extreme, 50 mm and 150 mm. Take at least 20 repeats, report the spread, then repeat the whole test after bumping the fixture. A system holding within ±0.2 mm on the artifact carries enough margin for the inspection.
Is a ToF camera ever the right choice at short range?
For volume, occupancy and safety sensing at 0.5 m and beyond, yes: direct ToF returns calibrated distance per pixel with millimetre-to-centimetre noise. Below roughly 0.2 m most modules report nothing, and their noise class dwarfs a millimetre anywhere they do report. Treat ToF as a logistics and presence technology, not a close-range metrology one.