Depth camera flat plane distortion: validate edge errors

A flat surface that curls upward near the image corners can indicate a range problem, a data-conversion mistake, or a persistent measurement error. A rendered surface alone cannot tell you which. This troubleshooting procedure is for depth-camera users who need to decide whether their measurements are usable, their setup needs adjustment, or their camera needs vendor evaluation. Start with documented operating limits, then measure plane residuals under controlled conditions before changing calibration.

Separate an unusual display from a measurement error

A tilted plane does not have constant depth across the image. Its measured depth should change with position, even when its physical surface is flat. A display that treats depth values as height can exaggerate that slope or make perspective effects difficult to interpret.

The useful question is not whether every pixel has the same value. It is whether reconstructed points lie close to a plane, within your application's measurement allowance.

Work through the following causes in order. This is an efficient test sequence, not a universal ranking of failure rates.

For the broader distinction between a genuinely curved reconstruction and a misleading fit, use the point-cloud plane curvature diagnostic. The procedure below focuses specifically on quantifying edge errors and collecting reproducible evidence.

Verify the depth data before testing the camera

Keep the original depth frame and its acquisition metadata. Record the stream mode, depth scale, coordinate convention, processing settings, and whether the image has been aligned to another stream.

A useful first check is to compare one unchanged frame in the manufacturer's supported viewer and your own application. Select several matching pixels and compare their numerical values. Check the center and corners, rather than relying on a single central sample.

Confirm these details before exporting:

• Preserve the native numeric representation instead of converting measurements into a display image.
• Apply the documented scale exactly once.
• Identify invalid samples using the documented validity rules.
• Exclude invalid samples from fitting rather than treating them as measured surfaces.
• Use intrinsics that belong to the actual depth stream and resolution.
• Determine whether the reported scalar represents axial depth or distance along a viewing ray.

That last distinction matters. A plane viewed along different rays does not generally have constant ray distance, even when it is perpendicular to the camera axis. Use the manufacturer's documented deprojection method to obtain camera-coordinate points.

If distances or clipping thresholds appear inconsistent, check the depth-scale and point-cloud cutoff workflow. Do not attempt to correct a unit error with calibration.

Check the closest surface against the documented range

The distance to the background is not necessarily the distance that determines whether the setup is valid. A raised object, an inclined target, or a projecting fixture may place part of the scene closer to the camera.

For a front-facing object whose height projects toward the camera, a simple axial estimate is:

For example, suppose the background is 320 mm from the documented measurement datum and the object projects 45 mm toward the camera:

Compare 275 mm with the device's documented operating limits. This calculation does not establish that any particular camera can measure there. For tilted objects, evaluate the closest corner or use the full geometry rather than subtracting nominal height alone.

Leave margin for placement variation, vibration, and the dimensions of the real part. Also check how the manufacturer defines distance: the housing face and the optical measurement datum may differ.

Move the test plane farther away while staying within the documented range, then repeat the same acquisition. Improvement with distance points toward an operating-condition problem, but does not by itself prove that calibration is correct.

The working-distance and positioning guide provides the related setup checks. Do not transfer a usable range from another camera, even if its images look similar.

Capture a repeatable flat-plane dataset

Use a rigid reference surface whose flatness is known well enough for the error you want to detect. An ordinary tabletop may be suitable for finding a large problem, but it is not automatically a dimensional reference.

Choose a diffuse surface and avoid glare, strong texture changes, and nearby objects during the first test. Keep both the camera and target stationary. Follow any documented stabilization procedure before recording.

1. Capture native depth first. Disable optional processing where supported, and record which operations cannot be disabled.
2. Fill the measurement region with the plane. Keep its physical boundary outside the evaluated area so that boundary artifacts do not dominate.
3. Record a short sequence. For example, save 30 frames to compare persistent shape errors with frame-to-frame variation.
4. Repeat at several distances. Use three positions within the documented range, including the intended application distance.
5. Repeat with a different target position. Translate the reference surface while retaining coverage of the evaluated image area.
6. Keep all settings comparable. If exposure or another setting must change, record it as a separate test condition.

Save raw data rather than only screenshots. A useful dataset includes timestamps, settings, the reference description, and the manufacturer's software and firmware versions.

Compare center and edge plane residuals

Reconstruct valid points in camera coordinates. Fit a plane to a predefined central region, then evaluate both that region and separate edge bands against the same plane.

For plane coefficients a, b, c, and d, the signed perpendicular residual is:

Coordinates and d must use consistent units. The residual sign depends on the chosen normal direction; keep that direction consistent between captures.

As a worked example, suppose the fitted plane is:

At x = 100 mm, the plane predicts z = 300 + 0.02 × 100 = 302 mm. If the measured point has z = 310 mm, its perpendicular residual is:

This calculation removes the plane's tilt. Subtracting a single central depth value would not do that.

Report more than the largest residual:

For example, an edge median that stays near +6 mm while individual points vary by roughly 1 mm suggests a persistent spatial effect. Random fluctuations centered near zero indicate a different problem.

A central fit establishes relative shape, not absolute distance accuracy. It can also hide error shared by the central region. Repeat the analysis using a whole-plane fit, and use an independently located reference plane when absolute accuracy matters. Keep region definitions fixed so that selecting different pixels does not manufacture an apparent improvement.

Isolate the target, camera, and processing contributions

A residual pattern becomes more useful when you know what it follows.

Translate the target across the field. If the abnormal region follows a particular physical patch, investigate target flatness or surface response. If it remains near the same image corner, investigate camera-coordinate effects, illumination geometry, and reconstruction.

Next, compare the native depth output with each processed stage. Alignment, hole filling, resampling, and smoothing can change edge appearance. Run the same residual analysis on each stage rather than judging only the final rendering.

Use these controlled comparisons:

• Native versus processed depth: identify whether curvature appears before or after processing.
• Continuous plane versus an object on a plane: isolate mixed-depth boundaries and occlusion.
• Fixed settings versus automatic settings: determine whether acquisition changes affect repeatability.
• Single active device versus the normal installation: where applicable, test whether other active illumination changes the result.
• Repeated captures versus one frame: separate persistent bias from isolated invalid measurements.

Do not assume that every corner error is ordinary lens distortion. Geometry depends on the depth principle, calibration, and processing pipeline. If incorrect rectification is suspected, use the separate depth rectification and distortion-coefficient explanation rather than experimenting with undocumented coefficient changes.

Decide whether to adjust, recalibrate, or request service

An edge region can have different uncertainty from the center. That does not make arbitrary curvature acceptable. Judge performance against documented conditions and the measurement tolerance your application requires.

If moving into the supported range removes the error, revise the mounting distance and verify the entire part envelope. If a target change removes it, validate representative production surfaces before accepting the arrangement.

If processing introduces the error, correct that stage and repeat the raw-data comparison. Cropping may be reasonable when the remaining region meets your coverage and accuracy requirements, but document the reduced usable area.

Temporal averaging can reduce some random variation. It does not remove a stable spatial bias merely because the output looks smoother. Likewise, fitting and subtracting an empirical curved surface is not automatically a valid calibration correction.

Use only manufacturer-documented calibration or recovery procedures. Obtain software and firmware from the manufacturer or its authorized distributors, and retain a record of changes. Do not overwrite calibration data with parameters from another unit.

Request vendor evaluation when repeatable residuals remain unacceptable after range, data interpretation, reference quality, and processing have been checked. Provide:

• Native frames and the settings needed to reproduce them.
• Reference-plane details and independently established distance information.
• Center and edge residual statistics at each tested distance.
• Valid-point fractions and temporal variation.
• A comparison between supported viewer output and application output.

Ask the vendor to assess the results against the device's documented performance conditions and advise on calibration, service, or replacement.

Choose replacement hardware from test evidence

Do not choose a replacement solely because another device produces a flatter-looking visualization. Compare candidates using the same reference, distance envelope, surface types, and acceptance metrics.

The 3D sensors category covers different measurement approaches. Use the camera finder to narrow the selection, then confirm the specific model's documented measuring range, field coverage, calibration procedure, and data interface.

For a projected structured-light candidate, review MV-DPS200P-02. Its suitability still needs to be assessed against your inspection requirements.

For another stereo approach, review MV-DLS300P-04. Do not assume that different stereo implementations share operating limits or identical behavior on your target.

Neither listing establishes an accuracy allowance for your installation. Request the relevant documentation and, where necessary, an application evaluation before committing to a replacement.

Verify the fix before returning to measurement

Repeat the original test after making one documented change. A convincing fix improves the numerical measurements without hiding failures through cropping, invalid-point removal, or stronger smoothing.

Use this acceptance checklist:

• Range: every required surface remains inside documented limits throughout placement variation.
• Reference: surface flatness is adequate for the error being evaluated.
• Data: units, validity rules, intrinsics, and coordinate conventions are recorded.
• Shape: center and edge residuals satisfy a predefined application allowance.
• Coverage: the valid-point fraction is acceptable in every required region.
• Repeatability: performance remains acceptable across repeated captures and operating distances.
• Traceability: raw frames, settings, and software versions are retained.

Do not accept a visually flat result if invalid coverage has increased or the dimensional error remains outside your allowance. Conversely, a dramatic rendering may be harmless if its vertical scale exaggerates small, acceptable residuals.

For help selecting a measurement approach, start with the camera configurator. If persistent edge errors remain unexplained, send your test conditions and residual measurements through application support before changing hardware.