3D visionTroubleshooting

Stereo depth noise on a flat plane: lighting and shadow fixes

Why a stereo depth camera reads patchy noise on a flat plane: shadow, tilt and texture causes, five quick tests, and the lighting and threshold fixes that hold.

LumaSpec Vision12 min read Share on LinkedIn

On this page 12 sections
  1. 01What the symptom looks like
  2. 02Cause 1: shadow bands starve the stereo match
  3. 03Cause 2: the plane is tilted, not noisy
  4. 04Cause 3: texture and ambient light
  5. 05Cause 4: a calibration bow
  6. 06Five captures that isolate the cause
  7. 07Fix 1: even fill light from the camera's side
  8. 08Fix 2: factory calibration reset, then re-verify
  9. 09Fix 3: clip the background with a depth threshold
  10. 10When light cannot fix it
  11. 11Prevention: keep the plane at one value
  12. 12Questions

You aim a stereo depth camera at an empty floor or wall and the flat plane refuses to read one value: stripes, dips and sunken patches where a single number should be. You need that background uniform so a depth threshold can lift the parts off it. The camera is almost never broken; the errors come from shadow bands, mechanical tilt, weak IR texture or a damaged calibration table, in that order of likelihood. This guide gives you the five-capture test that isolates the cause, then the three fixes that hold: even fill light from the camera's side, a factory calibration reset, and a threshold computed from your own measurements. The 3D sensors, bar lights and controllers that do the work are linked where each fits.

What the symptom looks like

On the depth colour view the plane shows as bands: pale where it reads close, dark where it falls away. The parts segment cleanly and the background between them does not, so a depth cut that works at one end of the frame fails at the other. Two error shapes tell different stories. Patchy stripes that move when the room light moves are stereo matching failures, and light fixes them. A smooth bow that stays put is geometry, tilt or calibration, and no amount of light will move it.

Cause 1: shadow bands starve the stereo match

An active stereo rig matches texture between two IR cameras, and its dot projector supplies texture on surfaces that have none of their own. A cast shadow does two things at once: it drops the reflected signal, and it draws hard edges between lit and dark patches, so the matcher locks onto wrong correspondences along those edges. The classic artifact is a sunken stripe: depth pulled off the plane, often by tens of millimetres at a metre-class mount, in a band that follows the shadow.

The test takes seconds. Capture the RGB view and the depth view of the same empty plane, then wave a hand or move a lamp. If the error bands slide with the shadow edges, this is your cause. The same mechanism fools any brightness-based check, which is why shadows fool vision checks across so many inspection types, and off-center parts spread the same geometry across the whole scene, as covered in edge shadows from off-center parts.

Cause 2: the plane is tilted, not noisy

A tilted plane produces a smooth depth ramp that the colour view stretches into what looks like noise. The arithmetic is plain trigonometry: across a width W tilted by angle α, depth swings by W × sin(α). On a 1.00 m wide floor at 3 degrees, that is 1.00 × 0.052 = 0.052 m of depth swing, more than a 50 mm part causes. No threshold can separate a part from a background that moves 52 mm on its own.

Read the raw depth at the left edge and the right edge of the empty plane. A smooth difference means tilt, and the fix is mechanical. For 50 mm parts on a 1.00 m scene, level the mount until the left-to-right difference stays under 10 mm, which is a tilt of 0.5 degrees or less: 1.00 × sin(0.5 degrees) = 0.009 m. Shim the bracket first; every later fix gets easier.

Cause 3: texture and ambient light

Two lighting failures produce the same patchiness with no shadow anywhere. Too little usable texture: a dark, matte floor absorbs the projector's dots, and beyond the projector's useful range the dots are too faint to match, so the depth image goes patchy or black — the failure behind black depth on textureless surfaces. Too much ambient: direct sun washes out the dot pattern, and autoexposure climbs until the IR image sits on its noise floor. That second failure and its fixes are covered in depth camera noise in bright ambient light and the autoexposure floor on IR depth.

The decision is one line: when sun is the cause, shade or relocate. Adding LED power protects the wrong thing, because what you are defending is projector contrast, not raw brightness.

Cause 4: a calibration bow

If the plane reads as a dome or cylinder that stays fixed relative to the camera through every lighting change, suspect the calibration table, often after a user calibration went wrong or a firmware change. Reset to the factory calibration with the vendor's documented procedure, then capture once more: the bow should vanish, and whatever spread remains is the ripple your threshold works with. What residual error a healthy sensor still carries, and how to quantify it on a flat plate, is measured in flat plane distortion edge error; the same sweep doubles as infield trueness verification after any correction change.

Five captures that isolate the cause

Run these in order and the cause names itself:

  1. Empty plane under the existing light: capture RGB and depth, and mark where the bands sit.
  2. Repeat with shades drawn, then with room light full: bands that move with the light are shadow bands, Cause 1.
  3. Read raw depth at both ends of the empty plane: a smooth difference is tilt, Cause 2 — compute width × sin(angle) before touching any setting.
  4. Add fill light on the shadow side and recapture: the patch spread should collapse to a few millimetres. Record the plane's minimum and maximum depth; that spread S is a number your threshold needs.
  5. If a bow survives steps 2 to 4 unchanged, run the factory calibration reset and capture once more.

After step 4 or 5 you hold the two numbers the fix runs on: mount depth H and floor spread S.

Fix 1: even fill light from the camera's side

Fill light works when it is wide, soft and steady. Mount it on the same side as the camera, high enough to clear the tallest part, and long enough to span the scene, so the plane sees one illumination field instead of pools and gaps. A diffused panel beats a row of point sources here: soft edges do not redraw shadow lines on the plane. Place it as close as the field of view allows, because illuminance on the plane rises fast as distance drops, and avoid grazing angles that stretch every part into a fresh shadow. The light-distance-versus-shadow geometry is worked through for photometric stereo lighting.

If visible fill does not move the depth numbers, the IR cameras are filtering it out; match the fill to the camera's own IR band instead. The bar lights category filters infrared at 850 nm and 940 nm, so the band can be matched directly, and the lights finder narrows by size and type. Two diffused 24 V models cover most rigs: the LS-BAR-0295 at 1272 × 22 mm for wide scenes and the LS-BAR-1276 at 372 × 28 mm for compact ones:

Bar lights1272 x 22 mm Diffused bar light (LS-BAR-0295)LS-BAR-0295
Size
400 mm +
Type
Diffused
$1,509.00View part
Bar lightsBar light, 372 x 28 mm, White (LS-BAR-1276)LS-BAR-1276
Size
200-400 mm
Type
Diffused
$439.00View part

Drive the light from a light controller rather than a raw supply so brightness holds frame to frame and mains flicker never reaches the image; flicker versus exposure is its own diagnosis, covered in black bars from flicker versus exposure. For steady continuous work the LS-CTL-0341 gives 4 channels, 8.3 A total output and constant-voltage 24 V, digitally controlled:

Light controllersDigital power supply controller, 4 ch, 8.3 A total output (LS-CTL-0341)LS-CTL-0341
Output
Constant voltage 24 V
Channels
4 channels
Control
Digital
$429.00View partBuild a system with it

For short exposures, a strobe-capable digital box such as the LS-CTL-0070 (4 channels, 200 W, RS232) pulses the light in sync with capture, which freezes motion and keeps room light out of the frame. Strobing is also how lighting acceptance numbers get written; the method is set out in ambient light rejection specification, with more options in the light controller finder.

Fix 2: factory calibration reset, then re-verify

Use the vendor's documented reset rather than hand-editing a table: the procedure restores the factory parameter sets together, and a mismatched pair reads worse than the bow you started with. Re-run capture step 4 afterwards and re-measure H and S, because the reset usually changes both. If the bow returns within days, the camera is being knocked or heated unevenly, and the fix is the mount, not the table.

Fix 3: clip the background with a depth threshold

With the plane lit and level, the last fix is arithmetic. Set the maximum depth of the threshold filter between the tallest part you keep and the shallowest depth the floor can read; every depth SDK ships this as a minimum/maximum distance filter, and setting it removes the background from the data stream itself.

d_part = H - h d_floor_low = H - S/2 D_max = (d_part + d_floor_low) / 2 h_min = H - D_max

Worked example, with a 50 mm part on a rig you have just levelled and lit:

  • H = 1.000 m, the raw depth to the empty floor at the mount point.
  • After the fill light, the plane reads 0.985 m to 1.015 m, so S = 1.015 − 0.985 = 0.030 m.
  • Shortest part to keep h = 0.050 m, so its top reads d_part = 1.000 − 0.050 = 0.950 m.
  • Shallowest floor reading: d_floor_low = 1.000 − 0.030/2 = 0.985 m.
  • Cut point: D_max = (0.950 + 0.985)/2 = 0.9675, set to 0.965 m.
  • Check both sides: floor pixels all read 0.985 m or deeper, so every one clips; the 50 mm part reads 0.950 m, 15 mm inside the cut, and the floor sits 20 mm outside it.
  • Smallest surviving height: h_min = 1.000 − 0.965 = 0.035 m. On this rig, parts under 35 mm need the next section, not a lower cut.

Recapture with a 50 mm gauge block as the acceptance test: it must survive, and the plane must vanish. Two leftovers belong in software. The part-background boundary carries a band of flying pixels a few pixels wide, so erode the mask back to the confident core, and drop pixels whose confidence value fails, the same masking move used on low-confidence depth.

When light cannot fix it

Three surfaces defeat lighting: matte black absorbs the dots, mirror finishes send them elsewhere, clear film ignores them. When the plane or the part is one of those, change the geometry or the sensor. Distance first: stereo noise grows with roughly the square of range,

delta_Z = (Z^2 × delta_disp) / (f × B) ripple_ratio = (Z_new / Z_old)^2

so moving the camera from 1.000 m to 0.600 m multiplies the ripple by (0.600/1.000)^2 = 0.36: the 30 mm spread becomes about 11 mm, and the same threshold logic then keeps parts down to roughly 20 mm with margin.

When the surface itself is the problem, a projected structured-light camera such as the MV-DPS200P-02 casts its own dense coded pattern from inside the sensor, so background texture no longer relies on room light:

3D sensorsHikrobot MV-DPS200P-02 Projected structured-light 3D cameraMV-DPS200P-02
Type
Stereo / structured light
$8,829.00 · quotedView and quote

For height-per-line work on a conveyor, a laser profiler measures each profile directly and skips stereo matching altogether; the 3D sensor finder filters both classes. Decision per case: parts 35 mm and taller at a 1.000 m mount, threshold alone; 20 to 35 mm, move closer or add IR fill on the matched band; smaller, black or shiny parts, change the sensor.

Prevention: keep the plane at one value

Once the rig passes, write the numbers down and watch them for drift:

  • Mount the camera rigidly; a bracket that rings after each pass adds its own depth noise, because mounting stillness during capture decides depth sharpness here as everywhere.
  • Level to 0.5 degrees or better on metre-class scenes, and re-check after maintenance.
  • Keep the fill light on a digital controller at fixed brightness, and repeat the empty-plane capture when lamps age or the sun's angle changes the shadows your fill is fighting.
  • Put the acceptance numbers in the test plan as a pass/fail line: at H = 1.000 m, spread S of 0.030 m or less and D_max = 0.965 m keep 50 mm parts separable with margin on both sides.
  • Log the empty plane's raw minimum and maximum each shift; drift in that pair is your earliest warning.

To spec the hardware end to end, start at the configurator's lighting step and match a sensor from the finders above. If the five captures do not settle your rig, send the RGB and depth pair with raw minimum and maximum through the contact page; the rest of the depth-camera fixes sit on the 3D imaging topic page.

Questions

Why does the depth of the floor change when I move around the rig?

Your body blocks the room light, so the shadow pattern on the plane moves with you, and the stereo matcher loses the match inside each new shadow band. Confirm it by comparing the RGB view with the depth bands: dark patches in RGB line up with sunken depth. Move the capture point so nobody stands between the light and the scene, or add fill light on the operator's opposite side, and the bands collapse.

How far should the fill light be from the floor?

As close as the field of view allows. Illuminance on the plane falls with the square of the distance, so halving the distance gives roughly four times the light on the plane, and a quarter of the distance gives about sixteen times. Keep the panel parallel to the plane, inside the shadow-casting objects' height, and long enough to span the scene. Then re-measure the empty-plane spread and set the threshold from the new numbers.

Can I remove the floor without adding any lights?

Yes, when the parts are tall relative to the floor's ripple. With a midpoint threshold placement the knife-edge limit is a part height of half the measured spread, and anything shorter clips away with the background. If parts are 50 mm and the empty floor reads within 30 mm, software alone works; if the spread grows toward the part height, thresholding eats the parts and the lighting or tilt fixes come first.

Does the camera's own IR projector cause the noise?

No, it is the texture the stereo match depends on, so never block or cover it. What harms it is competition: direct sun washes the dots out, dark surfaces absorb them, and glossy surrounds bounce them into odd paths. Keep the projector window clear, match any fill light to the camera's IR band instead of flooding the scene with white light, and confirm the depth image settles when you shade the scene.

How do I confirm the factory calibration reset helped?

Capture the empty plane before and after with identical lighting, and compare raw minimum and maximum depth across a fixed region. A calibration bow shifts both numbers and changes the shape of the plane's depth map, while ripple from shadows collapses once fill light is on. Keep the before-and-after captures with the rig paperwork so the next engineer can tell drift from damage, and repeat the sweep after any firmware change.