Somewhere in many depth-camera builds sits an infrared flood light that someone added 'for reach'. Sometimes it earns its place. More often it quietly raises depth noise or punches holes in the point cloud, and nobody connects the fault to the light. This guide gives you the mechanism in one formula, a worked example with the arithmetic shown, and a five-minute bench test that settles whether added IR helps or hurts your scene. It is written for anyone integrating speckle-projector and stereo depth cameras from the 3D sensors category with machine-vision lights, and it ends with a verdict for every scene you are likely to meet.
How a dot-projector depth camera builds depth
Most compact depth cameras are active stereo. A projector throws a fixed pattern of infrared dots onto the scene, each IR image is matched against a stored reference, dot displacement becomes disparity, and disparity becomes depth. The pattern is the texture. Surface paint and reflectivity matter only insofar as the dots stay visible, and if the pattern fades, the engine falls back on passive matching: it correlates whatever surface texture exists, which is far weaker.
Stereo precision follows one relation:
Z is distance, B the stereo baseline, f the focal length in pixels and σd the matching error in pixels. Worked example, with example inputs for a typical compact module: 1,280 px wide image, 87° horizontal field of view, 55 mm baseline, 0.2 px matching error.
f = (W / 2) / tan(FOV / 2) = 640 / tan(43.5°) = 640 / 0.949 ≈ 674 px
- At Z = 1,000 mm: σZ ≈ (1,000² × 0.2) / (55 × 674) = 200,000 / 37,070 ≈ 5.4 mm
- At Z = 2,000 mm: the Z² term quadruples it: ≈ 22 mm
- At Z = 4,000 mm: ≈ 86 mm
No amount of added light moves that curve. Light decides only whether the matcher can hold 0.2 px at all. That is why long-range stereo depth accuracy with a wide baseline starts from geometry, and why the 3D camera working distance and positioning guide treats working distance as the first setting, not the last.
Why added IR washes the pattern out
A flood adds a background level B on top of the dot signal S in every IR pixel, and the matcher can only use the contrast between them:
Two failure modes follow. Contrast falls wherever B approaches S, and correlation precision degrades with it. If bright surfaces clip near full scale, the dot amplitude is destroyed outright, and any auto-exposure reacts to the brighter scene by shortening the exposure, which starves the already-weak dots of photons.
The damage is distance-selective, because dot irradiance falls with the square of distance while the flood's contribution depends only on where you point it. Take a flood worth B = 10 gray levels against dots worth S = 100 at 1 m:
- At 1 m: C = 100 / 110 = 91%, the flood is invisible to the matcher
- At 2 m: S = 100 / 4 = 25, C = 25 / 35 = 71%
- At 4 m: S = 100 / 16 ≈ 6, C ≈ 6 / 16 = 38%, the pattern is more than half buried
A light that changed nothing at your 1 m bench distance can gut the pattern the day the camera moves out to 3-4 m. Up close it cannot help either: the dots already dominate, and all the added light can do is clip specular highlights.
On shiny parts it is worse still, because a broad flood paints large specular patches that both cameras see identically, and the matcher locks onto those shading edges instead of the dots. When the pattern fails for other reasons — fuzzy, black or thin surfaces that scatter or swallow the dots — the same weak passive fallback is all that remains, which the close-range depth dropout on fuzzy surfaces and thin parts article covers from the surface side.
When extra IR genuinely helps
Every case that works has one thing in common: the camera has little or no pattern to work with and would otherwise run on passive matching alone.
- Dark scene beyond the projector's reach. Past roughly 2 m the dot signal is down 4× or more against a 1 m working distance. A dim, diffuse, off-axis flood revives the passive signal, and this is the one case where the gain shows up reliably: at range, in the dark. At the bench it never does.
- Projector off or useless. Strong sun, glass and mesh scenes or hot specular surfaces can make the pattern itself the problem. The camera then runs purely passive, and a wavelength-matched flood becomes its only illumination; that build is covered in infrared flood illumination for stereo depth imaging.
- Dot dropout zones. Black, fuzzy or distant surfaces where the dots return nothing. A flood fills those holes with coarser passive depth. Expect higher noise there than on dot-based regions.
How much light does passive rescue need? Enough to be a real fraction of the light the scene already gets. Typical illuminance runs 100,000 lx in direct sun, 10,000 lx overcast and 300-500 lx in an LED-lit cell. A spot rated 500 cd (example rating) at 1.2 m from the target delivers 500 / 1.2² ≈ 350 lx, all of it in-band: decisive indoors at night, negligible outdoors.
Modern white LEDs emit almost nothing past 700 nm, so an LED-lit shop is effectively IR-dark; the same flood that does nothing by day transforms the IR scene at night. Under halogen or sun, already bright in the near IR, skip it. The ambient light rejection specification for vision acceptance tests shows how to put that comparison into acceptance numbers.
A ToF camera measures light differently, and the rule changes with it. The MV-DT01WMU is a ToF sensing camera: it reads the travel time of its own modulated emission, so an unmodulated flood adds background that lowers its modulation contrast. Confirm which technology you have before buying any light; the industrial 3D camera selection guide sorts stereo, ToF and structured light apart.
Wavelength, placement and drive
Match the wavelength to the camera, not to the LED bin on the shelf. A camera built around an 850 nm pattern and bandpass rejects a 940 nm flood almost entirely, and the reverse. The camera datasheet states its illumination wavelength, and the filter side of the pairing is covered in IR projector and depth camera: when the IR-cut filter blocks your plan. The spot lights category filters down to 44 models at 850 nm and 39 at 940 nm, so a wavelength-matched light is a catalog pick, not a custom build.
Small fields and coaxial lenses
191 models · from $109.00Placement decides most of the outcome:
- Mount the light 0.5-1 m to the side of the camera, aimed across the scene at 30-60° incidence. Off-axis light converts flatness into shading gradients, the texture passive stereo feeds on. A light beside the lens lights both images identically and adds nothing to correlate.
- Use a diffuse optic, not a tight spot. Hot spots create clip zones and shared glare.
- Start at 10-25% drive and raise it until the passive signal appears while the dots still stand clear in the raw IR image. A lighting controller turns that into a dial; the five reasons to use a lighting controller article lists what current control buys, and the lighting configurator walks the pick.
- Drive it steady, or pulse it in sync with the exposure. PWM that beats against the exposure time shows up as banding correlated across frames; pulsed IR illumination timing for low-light image capture works the timing arithmetic.
The five-minute test that settles it
Judge depth, never the IR preview.
- Aim the camera at a matte white target at the real working distance, projector and exposure as they will ship.
- Capture 200 depth frames with the flood off. Record the standard deviation of Z across a 100 × 100 px patch and the count of invalid pixels.
- Switch the flood on at its planned position and drive, let the exposure settle, and repeat the capture.
- Compare. Keep the flood only if depth noise drops by 10% or more, or invalid pixels fall materially, while the raw IR image still shows dot peaks standing at least 10% of full scale above the background.
- Repeat at the farthest working distance you will ship. The flood that was harmless at 1 m can be the one that buries the pattern at 4 m.
If the dots wash out, cut the drive before you move the light. If noise never improves, remove the light: the projector was doing the work. When the depth stream stays noisy with the lighting fixed, the stereo depth jitter troubleshooting guide applies the same measurement discipline from the sensor side.
Verdicts by scene
| Scene | What the flood does | Verdict |
|---|---|---|
| LED-lit cell, part at 0.3-1.5 m, projector on | Adds background only; the dots already dominate | Leave it off |
| Dark area, target at 2-4 m, projector on | Rescues passive matching as the dots fade | The one case that usually pays; test it |
| Projector off or useless (sun, glass, mesh) | The only illumination the camera has | Fit it, off-axis and diffuse |
| Daylight or halogen ambient | A drop in the bucket against thousands of lx of ambient IR | Skip it |
| Shiny metal at any distance | Shared specular patches beat the dots in the matcher | Avoid it; shape or mask the light |
| ToF camera | Constant background lowers modulation contrast | Skip it |
For an enclosed cell at a fixed working distance, the common industrial case, pattern strength beats added light every time. Pick the geometry first: the MV-DPS200P-02 projects a structured-light pattern for fixed-station inspection, and the MV-DLS300P-04 pairs stereo matching with a steered laser for larger scenes. A flood is a patch for dark or passive scenes, never a plan for reach.
- Type
- Stereo / structured light
- Type
- Stereo / structured light
Checks before you order the light
- Wavelength match. Read the camera's illumination wavelength from its datasheet and buy the same nm; a mismatched bandpass silently rejects most of the light you paid for.
- Diffuse optic and offset. Mounting 0.5-1 m off axis with a diffuser is what turns a flood into texture; a bare spot near the lens recreates the wash-out.
- Current control. A controller with an analog or PWM input lets you hold the flood under the wash-out threshold instead of guessing once at full power.
- Baffles and rating. Shield the emitter so neither lens nor eye sees the bare die, and rate the station against IEC 62471 at the working distance.
- An acceptance number. Write the 10% noise-reduction rule into the buy-off and measure it with the five-minute test. A light that cannot clear that bar does not ship.
When the usual advice is wrong
- 'Brighter IR preview means better depth.' The preview says nothing about depth noise; judge the depth map.
- 'More IR extends the range.' The Z² curve rules reach. Light rescues dropout; it does not move precision.
- 'Mount the flood right beside the camera to boost reach.' Near-axis mounting adds shared glare and coincident shadows; the offset position is what produces usable texture.
Set the scene up with the camera finder and light finder, run the five-minute test, and keep only a light that lowers depth noise. The 3D topic page collects the working-distance and selection guides, and you can contact the applications team with your working distance and target material if the geometry-versus-light decision is still open.
Questions
Does adding IR light improve a depth camera's accuracy?
No. Geometry sets the accuracy floor: stereo depth error grows with the square of distance, and no light changes that curve. Light decides only whether the matcher can hold its best subpixel error. Added IR improves depth only where the pattern has faded and passive signal was starving, typically a dark scene past about 2 m. Everywhere else it adds background that reduces dot contrast.
Why do my projected dots disappear when I switch on the IR flood?
The flood raises the background level in every IR pixel, so dot contrast falls; if bright surfaces clip, the dot amplitude is gone outright, and auto-exposure may shorten the exposure and dim the dots further. Cut the drive first, then move the light off-axis and add a diffuser. In the raw IR image, dot peaks should stand at least 10% of full scale above the background.
Should I use 850 nm or 940 nm light with my depth camera?
Use the wavelength the camera is built around; its datasheet states it. A camera with an 850 nm bandpass rejects most 940 nm light, and the reverse. Outdoors, 940 nm systems see less solar background because that wavelength sits in a water-vapor absorption dip, but the advantage only exists when the camera's optics were designed for 940 nm.
How far should an IR fill light sit from the depth camera?
Place it 0.5-1 m to the side of the camera, aimed across the scene at 30-60° incidence, behind a diffuse optic. The offset converts flat surfaces into shading gradients, which is the texture passive stereo matches on. A light mounted beside the lens lights both images identically and gives the matcher nothing new to correlate.
Can I add IR light to a ToF depth camera?
Generally no. A ToF camera measures the travel time of its own modulated emission, and an unmodulated flood adds constant background that lowers modulation contrast. The only version worth testing is a pulse synchronized to the camera's exposure, at long range in the dark, and then judged by the same depth-noise comparison you would run for a stereo camera.