Your stereo rig reports depth, but the numbers drift, and when you try to recalibrate, the checkerboard is found by the RGB camera and the right sensor while the left sensor reports nothing. That pattern almost always means the two eyes are not looking at the same picture. This guide walks the failure in the order we most often see at the bench: a screen used as the target, a dirty or out-of-focus lens on the one failing camera, and auto-exposure that set a different shutter on each sensor. You will confirm each cause from the raw frames, fix it, and only then rerun the calibration, because feeding barely-detected frames to the solver makes depth worse, not better.
One good eye, one failing eye: what it means
A stereo calibration tool asks one thing of every capture: the full board, found cleanly, in both grayscale eyes at the same instant. On these rigs depth comes from the disparity between the two mono sensors; the RGB sensor only paints colour onto the cloud afterwards. RGB passing the check tells you the board prints well and sits in view, and nothing more about depth. If the right eye finds the grid and the left does not, the solver is innocent and the left image is the variable.
The corner finder needs three things from each image: squares large enough, edges with real contrast, and a grid not broken by glare, banding or blur. Each cause below knocks out one of the three. Test in this order; the first explains most cases.
First test: save the raw frame from every sensor
- Set the target square to the camera axis at your working distance, however you hold it today.
- Capture one frame from each sensor at the same instant with the capture tool your SDK provides, and save the images exactly as delivered, before any extra sharpening or colour work.
- Open the left and right frames side by side and compare four things: overall brightness, stripes or bands, a glare patch over the squares, and edge sharpness on two squares of the same row.
- Measure one square's width in pixels on each frame; both readings should land in the 25-40 px band this guide uses.
If the failing sensor never streams at all, the fault is one layer down — device detection, cable or driver — and the USB camera not-detected troubleshooting guide covers that case; corner detection is not your problem yet.
Identical-looking frames that still fail point at target size or a broken grid. A left frame that is darker, brighter, banded or soft against a clean right frame points at one of the three causes below, and a left frame that passes only now and then is the same story at lower amplitude.
Cause one: a screen is not a calibration target
A display hands each eye a different target, through three mechanisms at once.
Refresh: a 60 Hz panel rewrites itself every 16.7 ms, one band of rows at a time. A short exposure catches the screen mid-refresh, so part of the sensor sees the new frame and part sees the old, and where that seam falls depends on when each eye's exposure starts. The left frame can carry stripes the right frame never shows. Backlights that dim by pulse-width modulation add whole-frame flicker on top.
Moiré: the panel's own pixel grid beats against the sensor's photosite grid and throws ghost squares that shift with a few millimetres of angle or distance. The two eyes sit tens of millimetres apart, so each sees its own ghost grid printed over the same board.
Glare: a glossy panel mirrors the room lights, and the mirror patch lands somewhere different from each viewpoint.
Print the target instead: matte paper or card on a rigid flat backing, no gloss lamination. Print at 100 percent scale, then measure three squares with a ruler and average the readings, because printer scaling quietly resizes files; the print-and-measure discipline in measured marker size in hand-eye calibration is the same habit. If corners still fail on a printed board with clean frames, inspect the grid for broken lines and thin ink, which checkerboard corners not detected: fix gaps in your target walks through.
How big the squares must be, with the arithmetic
Working rule: each square should span roughly 25-40 px in the frame where you hold the board. Below about 25 px the edge is only two or three pixels wide and blur or antialiasing erases the corner; much above 40 px and only a few squares fit in view, so every pose carries less information.
Worked example: a stereo camera with 1280 px wide mono frames and a 90° horizontal field of view, board held at 500 mm.
- frame width: 2 × 500 mm × tan(45°) = 1000 mm
- image scale: 1280 px / 1000 mm = 1.28 px/mm
- square: 25 mm × 1.28 px/mm = 32 px, mid-band
- board check: a 9 × 6 board of 25 mm squares spans 225 mm × 1.28 px/mm = 288 px, leaving margin for tilt inside a 1280 px frame
For that camera at half a metre, print 25 mm squares. At a metre the same camera needs about twice the square size; the checkerboard size for stereo calibration at 1 to 1.5 m guide runs that range, and checkerboard calibration fails at distance: depth camera target size shows how the pixel budget shrinks with distance. The fuller treatment of square and board sizing is in checkerboard calibration target size: how many squares and how big.
Keep the board fully inside both frames at every pose: a grid cut by one eye's frame edge is a dropped pair, not a partial success.
Cause two: the failing eye's optics
If the left frame is soft where the right is crisp, the target may be perfect and the eye is not. Causes in order of likelihood: film or dirt on the lens window, focus set at a different distance than yours, a knock that shifted one lens, and condensation after a cold-to-warm move.
The test takes two shots. Aim both sensors at the board at the working distance and compare edge contrast on the same square. Then aim both at a matte white wall and look for smears or shaded patches that shift when you tap the housing lightly — those are dirt, not focus.
Fix what the findings show. Clean the window as the manufacturer's instructions describe and re-shoot the raw frames. If your model exposes a focus adjustment, set it at the working distance and lock it; most stereo depth cameras are focus-fixed, so a left eye that stays soft after cleaning while the right stays sharp is a unit fault. The chessboard check in one stereo camera out of focus: diagnose it with a chessboard separates focus from damage in a couple of minutes. A soft eye cannot be calibrated into usefulness; replace the unit.
Cause three: each eye picked its own exposure
Auto-exposure loops run per sensor, and a checkerboard is half white and half black — close to the worst case. One eye can settle on the dark squares while the other chases the bright field and overexposes it. Saturation is the quiet corner-killer: blown white squares grow into each other, the grid lines between them thin to nothing, and the finder reports no board on a frame that looks fine to your eye.
Read the exposure and gain values your SDK reports alongside each stream. If the two mono sensors disagree, that disagreement is the smoking gun.
Switch both sensors to manual and set the same values, chosen from the dimmer eye's auto reading plus headroom. For example, if the dimmer eye settles at 12 ms, set both to 16 ms and confirm on the histogram that the white squares stay short of saturation. Keep gain equal too. If you must hand-hold the board, prefer a shorter exposure with more gain over a longer one that smears the grid while your hand moves. The per-sensor procedure in set auto exposure per stereo camera in a depth rig is exactly this step.
Recapture and recalibrate the pair
With both eyes detecting cleanly, rebuild the calibration from still, matched frames. The fixture is modest: a rigid camera stand, a flat backing for the board, and room to move the board through the poses.
- Mount the camera on the stand and the board on its backing. A rig that cannot move cannot blur: the LS-MNT-0040 gives a 700 mm working height on a 400 x 450 mm base, with a slide for the final distance trim, and leaves both hands free for the target.
- Type
- Camera stands
- Collect 15-30 pairs: board centred, then at each corner of the area both eyes share, then tilted 20-45° in four directions, then near and far inside the working band. The capture pattern is laid out in stereo depth camera calibration: target capture and stream setup.
- Hold every pose dead still through the exposure. Hand-carrying tilted poses is where wobble creeps in, so park the board on a stage instead: the LS-MNT-0001 tilt stage holds the printed board at a set slope between shots.
- Type
- CCD camera stand and positioning stage (LS-MNT-S01)
- Run corner detection over every frame before solving and drop any pair where either eye failed. Ten clean pairs beat thirty marginal ones.
- Solve and read the RMS reprojection error the tool reports — the average corner position error in pixels. With sub-pixel corner detection a healthy stereo solve lands well under one pixel; several pixels means the input frames, not the solver, need work. The stillness rule behind step 3 is spelled out in 3D camera mounting: stillness during capture decides image quality.
When depth is still bad afterwards
Detection clean on both eyes and an RMS error under a pixel, yet the cloud still swims: the runtime is the suspect now, not the calibration. Production stereo modules ship calibrated from better fixtures than a bench can offer, so recalibration is for a damaged or drifted calibration, not a tune-up for a working one. If depth degraded after a knock or a hot-cold cycle, the recalibration was the right call; if it was always mediocre, work on the scene instead.
Texture comes first: the matching stage needs detail to correlate, and untextured plastic, bare metal and black surfaces give it nothing at any calibration quality. Runtime exposure comes second: auto-exposure can reintroduce the left-right imbalance you just removed. Prove the rig on a matte, textured plate at a known distance and read the spread of the depth values against it; poor stereo depth quality: fix calibration and focus mismatch first continues from there, and 3D camera infield correction: verify dimension trueness on the line covers the measurement-grade check.
The decisions, in one place
- Screen as target: replace it with a printed matte board, squares measured with a ruler, size set by the arithmetic above.
- One soft eye: clean per the maker's instructions, then recheck; a still-soft eye means the unit is faulty.
- Exposure mismatch between eyes: manual mode, equal exposure and gain on both sensors, chosen from the dimmer eye.
- Detection now passes on both: recapture 15-30 still pairs and solve; expect an RMS reprojection error under about a pixel.
- Left eye still dark, banded or soft after all of the above: the sensor is faulty, and the fix is a replacement camera, not more recalibration.
A replacement drops straight into this same bench routine: a projected structured-light 3D camera such as the MV-DPS200P-02 calibrates from the same printed target, fixed exposure and stand-mounted poses.
- Type
- Stereo / structured light
To pick the camera or the fixture, start at the 3D sensors category and its camera finder, which filter by type and series; the bench side starts at the stand configurator. If a failure resists these checks, the contact page is the way to the applications team — bring your working distance and the raw frames from both eyes.
Questions
Can I use a screen as a stereo calibration target if I have no printer?
Treat it as a last resort. Set the panel to full brightness with adaptive dimming off, show pure black and white squares, and give both sensors the same manual exposure shorter than one display refresh, about 10 ms at a 60 Hz panel. Keep the screen square to the camera so the glare patch does not move between shots, and print a matte board as soon as you can: measure its squares with a ruler before you trust any recalibration.
How big should the checkerboard squares be at one metre working distance?
Run the same arithmetic as in the guide. On the example camera, 1280 px wide with a 90° field of view, one metre covers 2 × 1000 mm × tan(45°) = 2000 mm, so the scale is 1280 px / 2000 mm = 0.64 px/mm. Squares between 39 and 63 mm land in the 25-40 px band, and a 50 mm square sits at a comfortable 32 px. Swap in your own frame width and field of view for your numbers.
How many image pairs does a stereo recalibration need?
Plan on 15-30 good pairs. Cover the centre and all four corners of the overlap area, add near and far positions inside the working band, and tilt the board through roughly 20-45 degrees in at least four directions. More poses help only if each one is still and detected in both eyes; ten crisp pairs beat thirty where the board moved during the shot.
Can I recalibrate only the left camera of a stereo pair?
No. A stereo calibration estimates the two cameras' intrinsics plus the rotation and translation between them in one solve, so a saved calibration describes the pair, not either eye alone. If the left sensor returns clean, sharp, correctly exposed frames but the pair still misbehaves, rerun the full two-camera calibration. Replacing one eye's parameters by hand into an old result is not supported and usually makes depth worse.
Is reinstalling the driver or SDK worth trying when one eye fails?
Only if the failing sensor drops frames or will not enumerate. Corner detection runs on the delivered image, so a steady stream with failing detection points at the target, optics or exposure, which no reinstall changes. Before you reinstall anything, note each sensor's exposure and format settings, because a reinstall can reset them.


