Lighting for stereo depth accuracy: sunlight helps, glare hurts

Depth cameras read scenes differently indoors and outdoors, and light is the reason. The station that measures parcel volume cleanly in the yard opens holes in a dim hall, and a single glossy patch can blank a frame under a bare lamp. If you run a projector-based depth camera for volume measurement, bin picking or robot guidance, this gives you the matching physics end to end: what the stereo engine correlates, what the built-in IR projector contributes, how bright an added lamp must be in lux at the part, and where glare or IR oversaturation blanks the depth map. The lamp arithmetic is worked through for a 1,000 lm flood, and every scene type ends with a decision.

Light is texture: what a stereo depth engine matches

A stereo depth camera finds range by matching pixels between two views. It shifts a patch in the left image to its twin in the right one, converts the shift to depth through the known baseline, and repeats across the frame. Matching needs detail: fibres, print, edges, speckle. A smooth white wall gives the correlator nothing to lock onto, and those pixels come back as holes.

Active stereo adds a dot projector. Most projector-based depth cameras work in the near infrared around 850 nm, and their two imagers are band-matched to that wavelength. Where a scene has no texture, the projected dots create it, which is why a blank carton still returns depth indoors.

Sunlight then does something simple and powerful: it is broadband, so it carries near infrared too. Outdoors, every matte surface presents strong natural texture in exactly the band the imagers see, matching contrast rises, and depth noise falls. Standard illuminance tables put an office at 300-500 lx, an overcast sky at 1,000-10,000 lx and full daylight near 100,000 lx. A dim 200 lx bay sits about 500× below the sun, yet active stereo still works there — indoors, the projector rather than the room light supplies the texture the matcher needs.

What the IR projector does, and where it stops

The projector is a fixed-power texture source. Its dots spread over the scene, so pattern contrast on the target falls with distance roughly the way a lamp's output does: with the square of it. At the working distance the camera was designed around, dots dominate on featureless targets; well past it, they fade toward noise.

Three limits decide whether you need more light:

• Absorbing surfaces. Black foam, rubber and matte dark plastic return a few percent of incident light at any wavelength. Few dots come back and few scene features exist, so depth turns patchy no matter how bright the room is — the textureless-surface fix works this case through.
• Bright ambient. Under full sun the scene's own infrared texture is vastly stronger than the dot pattern, so the projector stops contributing. That is not a fault: sunlight has taken over its job, which is exactly why the same measurement reads better outdoors. Leave the projector on — it costs nothing in accuracy and takes over again the moment the rig moves into shade; the trade-offs sit in adding IR light to a depth camera.
• Retro-reflection. A shiny patch can bounce the projector's own energy straight back into the imagers and saturate them. Same mechanism as lamp glare, treated below.

A depth camera with no projector runs passive stereo only and lives or dies by scene texture; the passive stereo and laser class article spells out what that costs. For standoffs beyond the built-in reach, external pattern projectors with higher output work on the same principle and mount close beside the camera so their pattern lines up with the stereo baseline.

On parcel and volume stations the projector-and-camera pair often comes integrated: multi-DOE speckle structured-light cameras built for volume measurement combine the projector and the stereo pair in one housing, so the texture source is always aimed with the baseline.

Adding a lamp: which light helps the depth pair

Start from the benchmarks: about 200 lx on the part is a level at which active stereo already works on matte targets, and 100,000 lx of direct sun is the other working extreme. Everything added in between is about moving the part up from the first number with texture the imagers can actually see, without creating glints.

Band decides whether a lamp helps at all. Depth imagers band-pass around the projector wavelength, so the output of a phosphor white LED, concentrated below roughly 700 nm, barely registers on the stereo pair. A white flood brightens the room and the colour stream while adding almost no depth texture. Two choices genuinely move depth quality:

• An IR flood at the projector band. A dedicated 850 nm flood puts texture exactly where the imagers look, which is why it is the standard fix for dark, distant or absorbing targets. The IR flood for stereo depth guide works through the geometry.
• A genuinely broadband source. Halogen and other hot filaments carry real near-infrared energy, so they feed the depth pair as well as the colour camera. An ordinary LED flood still helps on visibly textured matte stock such as cardboard and wood, because those surfaces carry their own edges in every band.

Mounting beats wattage: set the lamp 30-45° off the camera axis, diffuse it over anything glossy, and judge the result by lux at the part rather than by lamp wattage. Lamp, driver and cabling can be pieced together from the lighting accessories finder or the extras configurator.

Worked example: a 1,000 lm flood over a volume station

Inputs, and where each comes from:

• Lamp flux Φ = 1,000 lm — a typical compact LED flood rating.
• Beam 60° × 60° full angle, half-angle θ = 30° — the usual optics on that class of flood.
• Mounting height d = 1 m above the part — a bench-scale volume station of the kind behind the indoor measurement question.
• Benchmarks 200 lx floor and 100,000 lx full sun — the working extremes set out above.

The rule in words: illuminance at the part equals lamp flux divided by the area the beam covers at that distance. A flat-faced beam footprint at distance d is a disc of radius d·tan θ.

At 1 m: tan 30° = 0.577, so the footprint diameter is 2 × 1 m × 0.577 = 1.15 m, the area is π × 0.577² m² = 1.05 m², and E = 1,000 lm ÷ 1.05 m² ≈ 950 lx — about 5× the 200 lx floor.

At 3 m: the footprint grows to π × 1.73² m² = 9.42 m², and E = 1,000 ÷ 9.42 ≈ 105 lx — below the floor.

Decision: hang this lamp about 1 m over the part and matte parcels get comfortable texture at roughly 950 lx. Under a 3 m ceiling the same lamp under-delivers at ~105 lx, so either drop it to about 2 m — E = 1,000 ÷ (π × (2 × 0.577)² m²) = 1,000 ÷ 4.18 ≈ 240 lx, just clear of the floor — or step up to a 2,000-3,000 lm unit, which returns 210-320 lx at 3 m. Real beams are centre-weighted, so treat these figures as ±30%. Past roughly 1,000 lx the accuracy gains flatten, and the next thing that breaks the depth map is not darkness but glare.

Glare: when good light blanks the depth map

A specular highlight reflects the source, not the scene. Shrink wrap, poly bags, tape and glossy labels turn a 100,000 lx sun into a highlight aimed down the lens; against an exposure set for a ~200 lx scene, those pixels receive about 500× the expected light, far beyond any sensor's headroom. They clip, the matcher loses its gradient, and the depth map drops exactly the region being measured.

Autoexposure extends the damage. With IR autoexposure enabled, one glint shortens the exposure for the whole frame and dims the projector returns everywhere else — the mechanism behind the autoexposure noise floor. Fixed bright ambient does the same more slowly, as the bright ambient noise fixes lay out.

Fixes, cheapest first:

1. Re-aim. Mount lamps 30-45° off the camera axis so specular reflections leave toward a wall, not the lens.
2. Enlarge the source. Diffusion — a diffuser plate, a light tent, a dome over small parts — spreads the same flux over a larger emitting area and drops highlight intensity.
3. Cross-polarize. A polarizer on the light and an analyser on the lens removes the specular component outright on shiny targets; the crossed polarizers method shows the set-up.
4. Tame fixed sun. When the sun sits fixed in frame, an ND filter on the imagers rebalances exposure; the ND filter sizing guide works the numbers.
5. Turn the projector down or off on texture-rich scenes where its return fights the ambient, where the camera's settings expose projector power.

For glare from fixed plant lighting rather than your own lamp, the same three moves — re-aim, diffuse, dim — are the substance of glare-reduction work, shown applied on a vision station in this glare reduction and lighting adjustment walkthrough.

When added light is the wrong answer

If the part is shiny, the tolerances are tight and the hall light changes hour by hour, stop adding lux and change the ranging principle. A laser profiler projects its own measured line and reads height from where that line lands on the surface: depth comes from triangulating a known pattern instead of hunting scene texture, and ambient light barely registers against a focused line.

One eye-safety rule comes with the class: the line is a classified laser, so respect the class label on the housing and never look straight into the line or its specular reflections.

Where scenes keep texture but plain stereo runs out of structure, projected structured light and galvanometer-scanned stereo cover the middle ground:

Once depth is trustworthy, volume is arithmetic on the height map — pixels to XYZ steps through the conversion — and the structured light vs stereo depth comparison sorts out which engine suits fine surface detail. Time-of-flight cameras solve the problem another way again, reading range per pixel from light travel time instead of texture.

Checks before you buy the light

• Match the band. Take the projector wavelength from the camera's own manual — 850 nm in most designs — and spec an IR flood to the same band.
• Budget in lux at the part. The 200 lx floor applies to matte targets; aim for 300-1,000 lx. Past that range, accuracy gains flatten while glare risk grows.
• Angle first. 30-45° off the camera axis, with diffusion over anything glossy.
• Plan the wiring. Floods mounted away from the driver need extension runs; LS-LAC-0048 cables come in 1, 3 and 5 m lengths with a 4-pin connector.
• Sync if you strobe. When colour and depth must see the same light state per part, sequence the lamp and the capture through a lighting controller rather than a free-running lamp.
• Read the confidence data. After each change, check the depth and confidence images: holes on glossy patches mean clipped glare, holes on matte patches mean missing texture, saturated blobs mean too much IR return.

The decision for each case

If a depth map still shows holes after these changes, describe the scene to us through the contact page — surfaces, standoff and the light you have now are enough to shortlist lamp and sensor together. To browse the sensor side, the 3D sensors category and the camera finder filter by type and range, and the 3D camera topic hub gathers the rest of the depth guides.