Depth camera noise in bright ambient light: what fixes it

A depth map that was clean on the bench turns to gravel the moment sunlight hits the scene: edges fuzz, holes spread across shiny or dark patches, and the part's silhouette wobbles frame to frame. The cause is almost never a broken camera. Bright ambient light buries the infrared dot pattern the camera projects, and the stereo match runs out of contrast to lock onto. This guide walks the likely causes in order, gives the exposure, filter and shading fixes with their numbers, and shows how to measure depth noise so you know the fix worked.

Why sunlight raises depth noise

A depth camera sees the scene by its own light. A projector on the camera throws a dense pattern of infrared dots or speckle across the part, two or more synchronised imagers view it, and the SDK computes depth from the shift between the views. Dim the pattern and the match has less to lock onto; drown it and depth noise climbs with every stop of added light.

Direct sun is the extreme case. A full sun delivers about 100,000 lux — an order of magnitude above an overcast sky and hundreds of times above indoor lighting — and silicon imagers respond across roughly 400 to 1000 nm, straight through the projector's infrared band. The pattern's contrast at the sensor falls as ambient rises; once a sun patch clips the pixels at full scale, the modulation is destroyed outright.

Distance compounds it:

Here f is the depth pair's focal length, B its baseline and d the disparity between views. A disparity error σ_d turns into a depth error that grows with the square of distance, and as a working estimate σ_d itself grows in proportion to 1/C, the inverse of pattern contrast. Halve the contrast and depth noise doubles; the same washout at double the distance makes it four times worse again. The lifted noise floor also shows in the raw IR image, and IR depth image noise in bright light explains how to read it.

Likely causes, in order of probability

Work down this table in order; the first row covers most daylight complaints.

The emitter on/off capture technique settles the first row in one pair of frames: pattern only, then pattern plus whatever the room contributes.

Turn auto-exposure off, then set exposure by hand

Depth SDKs enable auto-exposure at startup. The camera rebalances exposure, and sometimes gain, to hold an average image brightness — it balances the whole frame, not the pattern against the background, so in sunlight it merely re-aims the same washout at a target grey level. Enumeration output from a typical device shows the arrangement: an exposure option spanning 39 to 10000 µs with a default of 312 µs while auto-exposure is on, one option per stream.

Switch the auto-exposure option off before writing an exposure value; on many SDKs a master auto-exposure option must clear first, and the options error article shows the failure you get if it does not. Query each stream's exposure range from the API before you set, and mind the units the limit uses — microseconds on most, but not every, option.

Set the exposure so the brightest sunlit patch sits below about 80% of full scale; that keeps the pattern in the sensor's linear range with headroom as the patch moves across the part. On a moving line, cap exposure by part speed instead — the motion-blur cap article works that arithmetic. In a two-camera rig, set auto exposure per camera so one stream's loop does not fight the other's.

Shade, shroud or re-site the station

Blocking the direct beam costs a sheet of metal and an afternoon, and it attacks the dominant term. The sun's direct component carries most of its irradiance in a narrow beam; diffuse skylight spreads the remainder over the whole sky and is orders of magnitude weaker at the camera. A shroud — a matte-black tunnel around the field of view — removes the beam and leaves the weak diffuse component the pattern can survive. Mount the station so the sun never crosses the window's path during production hours, and curtain dock doors or trolley openings that open into daylight; a strip curtain over the seconds of capture is often enough.

Keep every surface inside the shroud matte. A glossy baffle turns one sun into a hundred moving highlights, which is worse than the sun itself.

Beat the sun with a narrowband filter

A narrowband filter rejects ambient where no software can reach. Match it to the projector, not the room: depth projectors commonly centre near 850 nm or 940 nm, and the camera's data sheet or option list names the emitter. Order a bandpass centred on that wavelength with a 10 nm full width and a peak transmission around 90%, mounted wherever the camera's optics accept filters. Two practical points: any protective window in front of the lens must transmit at the passband, and exposure will need a small rise because the filter also dims the pattern to its transmission.

Check the filter against a source before you mount it — the near-IR bandpass contrast test takes minutes. On 2D stations lit with red LEDs the same trick applies with visible-light filters; the red-LED bandpass article shows that variant.

Worked example: how much a 10 nm filter at 850 nm buys in direct sun

Inputs. Full sun under the standard AM1.5 spectrum delivers about 1000 W/m² in total, and roughly 650 W/m² of that falls between 400 and 1000 nm, the band a silicon imager responds to. At 850 nm the sun's spectral irradiance is about 0.12 W/(m²·nm). The projected dots sit at 850 nm, inside a 10 nm full-width filter with a peak transmission of about 0.9; both are typical catalogue values.

Rule. Ambient reaching the sensor equals spectral irradiance × filter bandwidth × transmission. The rejection factor is the broadband sunlight divided by that in-band remainder.

Arithmetic. In-band ambient: 0.12 W/(m²·nm) × 10 nm = 1.2 W/m², × 0.9 = 1.08 ≈ 1.1 W/m². Rejection: 650 W/m² ÷ 1.08 W/m² ≈ 600. The dots pass at 90%, so the pattern-to-background ratio improves by roughly 600 × 0.9 ≈ 540.

Decision. No exposure setting can match that, because exposure scales the dots and the ambient together. If the station holds depth indoors and loses it within seconds of direct sun, order the 10 nm filter at the projector's wavelength; a 600-fold background cut converts an outdoor washout back into an indoor noise budget. If noise stays high after filtering with the beam shaded, ambient is no longer the limiting term — move to projector power, a longer baseline, or a closer working distance.

Raise projector power inside the safety limit

Most depth SDKs expose the emitter's drive as a laser power option. Raising it lifts the dots against an ambient background the option cannot touch, and at short range a step or two of power is often the whole fix. Raise it until depth noise on the flat target stops falling, or until the option reaches its maximum, whichever comes first.

Near the camera's close limit, gain and exposure trade against each other and against depth noise; tune them together rather than one at a time.

Measure depth noise to confirm the fix

Diagnose and close the loop with one number: the standard deviation of depth on a flat matte target at the working distance. Capture 100 frames in production lighting — morning and noon if the station sees daylight — and compute the deviation over a central patch a few hundred pixels across. Before any fix, the number tells you which row of the causes table you are in; after each fix, the ratio tells you what the change bought.

Work to targets rather than a law of nature: hold depth noise under about 1% of distance for guidance and presence tasks, and under about 0.1% of distance where the camera feeds measurement. A bench structured-light camera indoors typically lands a few tenths of a millimetre at 0.5 m, so 1% of distance is a loose ceiling and 0.1% a demanding one. Precision and trueness are different specifications; the precision, trueness and spatial resolution article keeps the vocabulary straight when you compare cameras by this number.

When it is not the light

Three failure modes masquerade as ambient washout. First, calibration or focus mismatch between the imagers produces noise that ignores the lights entirely; a flat-target capture with the emitter on and room lights off isolates it, and the stereo calibration article walks the repair. Second, heat: range and precision sag as the camera warms, so an afternoon regression can be thermal rather than solar — the thermal derating article gives the tests. Third, the part itself: deep black or mirror-flat surfaces return too little pattern at any light level, and no filter will rescue them.

If the part is the problem and depth was only a route to surface inspection, control the light instead of the sun: the MV-PSA-12UM-C4D30-WDM pairs a 12 MP area-scan camera with a 300 mm white dome light, and its own illumination swamps the ambient so the sky stops mattering.

Choosing hardware for a bright site

When shading, filtering and power cannot hold the noise targets, the hardware has to match the lighting. Three families in the 3D sensors line cover the common bright-site cases:

The MV-DPS200P-02 is a projected structured-light 3D camera for close-range detail. The MV-DLS300P-04 is a laser galvanometer stereo camera for scanned-line work on larger parts. The MV-DT01WMU is a ToF sensing camera when a direct depth frame over the whole scene is enough. Whichever way you go, read how the data sheet states ambient-light tolerance before trusting an outdoor range claim; the ambient-light specification guide for outdoor laser distance sensors shows how that is done for one sensor family.

Shortlist with the camera finder, size the rest of the station at the configurator's camera step, and more repair guides live in the 3D camera guides. If the same station also runs 2D checks, beat ambient there by strobing: pulse the LED only during exposure through a light controller, as the ambient light control article lays out. Send the scene, its sun exposure and your noise targets with a request through the quote page and we will come back with a configuration before you order.