Minimum object size for 3D depth camera detection

A depth camera datasheet rarely answers the smallest-part question head-on. It lists a pixel count, a field of view and often a line such as object size 200 × 200 mm at 18% reflectivity. That looks far larger than the 26 mm part sitting on your bench. The line is not your limit, and neither is your segmentation code. Two physical numbers decide what you can detect: the lateral pixel pitch at your working distance and the depth noise on your surface. This guide works both out for a 350 mm standoff, shows arithmetic you can redo for any distance and any imager, and closes with the decision for small-part detection plus the cases where structured light or a laser profiler is the better tool.

What sets the smallest object a 3D depth camera can detect?

Two physical limits decide it, and neither is your detector: the object must span at least about 5 pixels across in the depth map, and any relief you depend on must rise more than 2-3 times the camera's depth noise above its surround.

A time-of-flight camera returns one depth value per pixel of its sensor array, so every limit traces back to that grid. Lateral detail is bounded by pixel pitch — how many millimetres of scene each pixel covers at the working distance. Depth detail is bounded by noise: the same pixel does not return the same distance twice, and relief shallower than that spread drowns in the surface. Your detector merges, thresholds and masks, but it cannot manufacture pixels the imager never captured.

Plan for at least 3-5 pixels across the object's narrowest dimension for a presence decision, 8-10 for a shape you must classify, and 20 or more for a clean instance mask. The pixel rule used to size cameras for defect detection carries over unchanged: count pixels across the feature, not pixels on the sensor.

How do you calculate the pixel pitch at your working distance?

Pixel pitch is field width divided by pixels across. Worked at the 350 mm standoff used throughout this guide, a 100k-pixel imager framed on a 200 × 200 mm field — about a 352 × 288 array — gives 200/352 ≈ 0.57 mm and 200/288 ≈ 0.69 mm. Plan on roughly 0.6 mm per pixel.

Worked example, the 26 mm part at that standoff. A 200 mm wide field at z = 350 mm needs FOV = 2 × arctan(100/350) = 31.9°, a narrow, telephoto-like field. Across a 352 × 288 imager the pitch is 200/352 = 0.57 mm horizontally and 200/288 = 0.69 mm vertically. The 26 mm diameter therefore spans 26/0.57 ≈ 46 pixels horizontally and 26/0.69 ≈ 38 pixels vertically — call it 40. The 29 mm part spans 42-51 pixels and the 66 mm length about 116, so every measure clears the 5-pixel floor by an order of magnitude.

Most industrial ToF cameras carry a wider field, near 60°. At the same standoff that covers 2 × 350 × tan(30°) = 404 mm, the pitch becomes 404/352 ≈ 1.15 mm, and the 26 mm part spans 26/1.15 ≈ 23 pixels — still detectable, with thinner shape detail. Cropping a region of interest in software recovers none of it: a crop discards pixels, it does not shrink the pitch.

If the SDK exposes intrinsics, object-space pitch is simply z/fx. An illustrative fx of 616 pixels — a 32° field on a 352-pixel-wide imager — gives 350/616 = 0.57 mm, matching the narrow-field case. The ToF camera intrinsic parameters guide shows where fx lives, depth camera pixel size walks the same field-of-view arithmetic, and pixel pitch without a datasheet shortcuts it.

Type your own field width and smallest feature into the resolution calculator:

Is that standoff too close for a ToF depth camera?

Most industrial ToF cameras measure from about 0.2-0.5 m outward. Near-field illumination and ambient rejection set that floor, so a 350 mm mount sits inside the band. It is the one number to confirm for the specific model before you drill the bracket.

Below the minimum range, close objects can return as speckle instead of geometry; the failure modes are catalogued in why close objects look like noise. Above that floor, closer is better on every axis: depth noise scales about with the square of distance and pitch scales linearly. At 175 mm the 32° field shrinks to 100 × 100 mm, the pitch halves to about 0.3 mm, and the 26 mm part spans 75-90 pixels, with the 66 mm part still clearing the field by 17 mm per side. Standoff and mount trade-offs are laid out in the 3D camera working distance guide.

The 3D sensors range includes ToF imagers of exactly this class:

That imager class is the right starting point here once the mount respects the minimum range. The pitch arithmetic above is what turns its pixel count into a yes-or-no answer.

What does the 200 × 200 mm object size spec actually mean?

That datasheet line rates the camera's built-in detection software. It is the smallest target the vendor's on-camera detector guarantees to flag at the stated reflectivity, and it is not the smallest object your own pipeline can find in the depth data.

Cameras aimed at safety and logistics publish the figure because their integrated detectors must catch rated targets — a person, a parcel, a vehicle — anywhere in the full field, without knowing in advance what the object is. Your application inverts that: you know the part, you frame the field on it and you run your own detector over every pixel. Your floor is the 5-pixel rule from the first section, which at 0.6 mm pitch is roughly a 3 mm feature — about 65 times smaller across than the 200 mm rated target. The same rated-target logic sits behind person detection with a depth camera. The 18% reflectivity half of the line belongs to the same guarantee, and it decides the next question.

How much does surface reflectivity change the answer?

Depth noise scales roughly with the inverse of returned signal amplitude, and amplitude scales with reflectance. So a part at the 18% floor returns about five times noisier depth than a 90% white target: 90/18 = 5.

Run the distance scaling first. A datasheet noise figure of 5 mm at 2 m, typical for this class, scales by the square of the distance ratio: 5 × (0.35/2)² = 5 × 0.031 ≈ 0.15 mm at the standoff on a bright target. Multiply by five for an 18% part and depth noise sits near 0.8 mm; a 6% dark rubber multiplies it by fifteen, to roughly 2 mm. Keep the relief you must see above 2-3 times those figures: about 0.4 mm of resolvable height on a bright part, about 2 mm at the floor — both far under the relief of bolt heads and caps.

Three counters work before you change hardware: give the imager more exposure time (the usual lever, traded against motion blur on moving parts), pick a higher-amplitude imager mode if the camera offers one, and lift the surface itself. Matte-white tape on the top face or a light bead-blast on the measured area returns more signal than any downstream filter. Where detail still drops out at edges and low-contrast patches, the causes match why a ToF point cloud loses detail on small distant objects.

Can a Mask R-CNN pipeline work on a part that spans 40 pixels?

Yes: instance segmentation wants roughly 20 or more pixels on the object's narrow side, and the 26 mm part spans about 40 at 0.6 mm pitch. The mask holds twice the pixel budget it needs.

Annotate at the pixel scale you will deploy. A mask boundary one pixel ragged is 0.6 mm of error on this set-up, and depth edges are ragged by 1-2 pixels anyway, because each boundary pixel mixes part and background and lands between the two depths — the mixed-pixel effect behind flying pixels at depth edges. Threshold on the camera's confidence or amplitude image so those mixed pixels drop out before the mask head sees them.

Feed the depth channel to the network even where 2D contrast would find the part: relief separates a real bolt head from a printed picture of one, and height features survive shading changes on the line. Keep capture and inference frames aligned; the hand-off is its own integration step, covered in depth camera object detection: connect frames to inference. If the camera offers a reduced-resolution mode for speed, re-run the pitch math before accepting it — the mode is set per changing depth resolution at capture time.

When should you switch from ToF to structured light or a laser profiler?

Change technology when the feature you must resolve closes on the pixel pitch: below roughly 3 mm lateral at this standoff, or when depth detail must be finer than about half a millimetre. Structured light or a laser profile sensor is then the better tool.

Projected structured-light cameras trade some speed for density: they flood the scene with a pattern and recover a dense surface per frame, which pays on small parts that hold still. Laser profile sensors scan one line at a time and rebuild the surface from part motion, with depth resolution in the micrometre class and sub-millimetre point spacing at close range — typical figures for the class, and the reason they dominate close-tolerance checks. The downstream pipeline stays simple if you feed the dense result as a 2.5D height map rather than raw points.

Both classes are stocked in the same 3D sensor range, matched to different jobs:

The projected structured-light camera suits dense surfaces on parts that hold still between frames. The laser profile sensor suits parts that move, where depth precision decides pass or fail.

What should you check before you order the camera?

Five numbers decide the order: pixel pitch at your standoff (0.6 mm here), pixels across the part (about 40), the camera's minimum range (0.2-0.5 m typical for ToF — confirm it for the model), the part's reflectivity (stay at or above the 18% floor), and the relief you must see against 2-3 times the depth noise (about 0.4 mm here on a bright part).

Run down the list before you commit to a mount:

• Pixel pitch at the standoff. Field width ÷ horizontal pixels: 200 ÷ 352 ≈ 0.57 mm here. Every lateral number downstream comes from it.
• Pixels across the part. 26 ÷ 0.6 ≈ 40 pixels; hold at least 5 for presence and 20 or more for clean masks.
• Minimum range. ToF class typically 0.2-0.5 m; a mount under it turns the near field into speckle instead of geometry.
• Reflectivity of the real part. Rated performance ends at 18%; a 6% dark surface multiplies depth noise up to about 15 times against a white target.
• Relief against depth noise. Scale the datasheet noise by (z1/z2)²; at the standoff that lands sub-millimetre on a bright target, so 1 mm-class features resolve.
• Field coverage at the closest mount. At 175 mm the 32° field is 100 × 100 mm, which clears the 66 mm part with 17 mm per side.

Decision for the parts in question: both the 66 × 29 mm and the 52 × 26 mm cylinders are comfortably detectable from the 350 mm standoff on this imager class — about 40 pixels across at 0.6 mm pitch, with depth noise an order below their relief — as long as the mount stays above the minimum range and the surface returns at least 18%. Below roughly 3 mm features, move to structured light or a laser profiler. Size the hardware from the camera finder or start at the camera configurator; if the part also needs print or surface checks, an area scan camera covers the 2D channel beside it. The 3D vision topic page carries the surrounding integration guides, and contact reaches the applications team for a second pass at the math.