Close range depth dropout on fuzzy surfaces and thin parts

A fluffy surface can look well exposed in a color image while producing very little usable depth. A narrow handle may appear in one frame and disappear in the next. Neither symptom proves that a software filter removed good measurements. This troubleshooting sequence separates viewing geometry, correspondence failure, output conversion and optional filtering. You will establish a controlled baseline, compare supported outputs and measure whether each adjustment preserves real detail rather than merely filling holes.

1. Separate missing measurements from display problems

Start with a stationary camera and target. Save the native depth frame, its validity mask if available, and the matching camera images supported by the device. Record the stream profile, exposure settings, processing configuration and timestamps.

Inspect numerical values, not just a colorized preview. A dark preview region can represent invalid data, valid distance outside the display range, or a rendering choice. Read the documented invalid-value convention before counting missing pixels. Do not assume every device uses zero.

Use this order to isolate the cause before adjusting filters:

The stereo disparity troubleshooting guide covers broader correspondence faults. This procedure concentrates on close working distances and fragile surface detail.

2. Test working distance and visibility first

Check the documented operating range for the selected acquisition mode. A housing-to-target measurement may differ from the manufacturer's distance reference, so use the specified datum.

A target can sit inside the apparent overlap of the camera views yet remain outside the supported matching range. At close distance, the required disparity becomes larger. One camera may also see a surface patch that the other camera cannot see behind a nearby edge.

Run a simple geometry test:

1. Place a matte, textured board beside the difficult target, approximately at the same distance.
2. Capture a stationary sequence with the original configuration.
3. Move both objects farther away without changing processing settings.
4. Repeat the capture and compare valid coverage on the board and target separately.
5. Rotate the difficult target slightly and observe whether missing regions move with its silhouette.

If both objects improve substantially with distance, investigate operating range and overlap before filtering. If only the fuzzy target fails, surface structure and image signal become stronger suspects.

A thin handle can expose background in one matching neighborhood and foreground in another. Individual fibers can also obscure neighboring fibers differently between views. More exposure cannot create a shared view where one does not exist.

If visible features do not align along the expected matching direction, check calibration and the rectified stereo video workflow before interpreting disparity differences.

3. Check disparity range and image sampling

For a rectified stereo pair with aligned principal points and a compatible disparity convention, the basic relationship is:

This describes ideal geometry, not a guarantee of valid correspondence. Calibration offsets and stored disparity scaling must be handled as documented.

Consider an illustrative pair with a focal length of 600 pixels and a baseline of 50 mm. At 400 mm, the expected disparity is:

Halving distance doubles the required disparity. Whether either value is supported depends on the device and mode; do not infer a search limit from this example.

Thin-feature sampling creates a separate constraint. If a 640-pixel image spans 320 mm at the target plane, the scale is approximately 2 pixels/mm. A feature 2 mm wide spans only 4 pixels. Its edges and matching neighborhood can include substantial background even though the feature is visible.

Use the camera resolution calculator below to explore feature sampling; pixel count alone does not predict depth completeness.

For calculations based on an aligned color image, first verify that its geometry matches the depth stream. The intrinsic parameter conversion guide explains why resizing and cropping change the relevant camera parameters.

4. Inspect the images used for matching

A useful color preview does not prove that the stereo matching images contain usable texture. Inspect the actual matching streams when the documented interface exposes them.

Look for clipped highlights, dark regions, blur and differences between the two views. Fuzzy material may contain many fine edges, but those edges are not necessarily stable or mutually visible. Similar-looking fibers can also produce ambiguous matches.

Adjust one acquisition variable at a time. Longer exposure may improve signal on a stationary target, but movement during exposure can blur the feature you need. Increased gain can brighten the image while amplifying noise. Neither change repairs occlusion.

Keep automatic adjustments from changing between comparison runs where documented controls allow this. Record actual exposure and gain rather than assuming requested settings were applied. The per-frame metadata guide explains that verification step.

For low-signal surfaces, follow the dark-surface stereo lighting checks. If an entire matching image is unexpectedly dark, use the infrared black-image diagnostic sequence first. Visible fill lighting may not improve an infrared matching stream.

5. Compare supported disparity and depth outputs

Use only stream modes, export options and processing controls documented for the installed software and firmware. Obtain software and documentation from the manufacturer or its authorized distributors.

Do not assume that an output named raw disparity contains every tentative match. It may already include device-side consistency checks or validity decisions. An application cannot reconstruct rejected internal candidates unless the device exposes them through a documented interface.

Build a comparison using these steps:

1. Confirm the documented availability of disparity, native depth, aligned depth and processing controls.
2. Save the original supported outputs before preview colorization or point-cloud conversion.
3. Record each output's dimensions, timestamp, units, scale and invalid-value convention.
4. Disable optional host-side filters individually where supported, saving the configuration after each change.
5. Compare corresponding frames in the same pixel coordinate system.
6. Repeat the original configuration to check that the scene has not changed.

If outputs cannot be acquired simultaneously, use a stationary scene and acknowledge that sequential captures are not identical measurements.

Valid disparity with invalid depth points toward conversion, range checks or processing, but only after accounting for geometry and scaling. Missing data in both outputs does not prove that no match was attempted; it only establishes what the exported interfaces provide.

Alignment can create a different hole pattern through reprojection and occlusion. Compare native outputs before comparing them with an aligned color view. If a stream cannot be decoded reliably, resolve the pixel-format and missing-stream checks before judging its content.

6. Measure coverage and flicker with a fixed test

Choose a fixed region on the target, excluding background where possible. Include an interior region and a narrow-edge region separately; their failure mechanisms can differ.

For each configuration, capture 100 stationary frames as an initial screening test. This is a practical comparison length, not an accuracy certification. Keep the camera, target, illumination and display scale unchanged.

Calculate coverage from the documented validity convention:

For example, 6,000 valid pixels in a 10,000-pixel target region give:

If another setting produces 8,000 valid pixels, coverage becomes 80%, an increase of 20 percentage points. That is useful only if the added pixels represent the target rather than background leakage or filled estimates.

A handle pixel valid in 40 of 100 frames has 40% availability. A later configuration giving 90 valid frames improves availability to 90%, but inspect its depth values and behavior during movement too.

For pixels that remain valid, report a spread such as the interquartile range or standard deviation. Keep validity changes separate from depth variation; averaging invalid codes into a distance statistic gives misleading results. Stability is also not accuracy: comparison against a known reference is a separate test.

7. Tune stability without erasing the feature

Once geometry and signal are acceptable, test optional processing in small, documented increments. Change only one setting per run and retain the baseline capture.

Control names and directions vary. A larger numerical setting does not universally mean stronger filtering or higher confidence, and confidence values are not automatically calibrated probabilities.

For a moving handle, test appearance, movement and removal. A stable-looking surface that remains after the object leaves is unsuitable for fresh position measurements.

When checking a reconstructed cloud, retain the depth validity mask. The missing XYZ point diagnostic guide helps separate depth loss from later conversion or visualization problems.

8. Decide whether hardware evaluation is necessary

Do not replace hardware solely because a filtered preview looks incomplete. First establish whether the application needs a continuous surface, a reliable central patch, a silhouette, or the position of a narrow feature. These are different acceptance criteria.

If the task still fails within documented operating limits, compare 3D sensors using the camera finder. Request evaluation on the actual fuzzy material and smallest relevant feature, not just a flat calibration target.

The MV-DLS300P-04 is listed as a laser galvanometer stereo camera. Its suitability requires application-specific evaluation.

The MV-DPS200P-02 is listed as a projected structured-light 3D camera. Treat it as another evaluation candidate, not a guaranteed remedy for missing depth.

For either candidate, confirm operating distance, smallest measurable feature, acquisition timing and documented output access before selection. The catalog descriptions alone do not establish these capabilities.

9. Save an acceptance checklist before ordering

Use a short record that another operator can repeat:

• Geometry: distance reference, target angle and common viewing region recorded.
• Acquisition: stream profile, exposure, gain and illumination saved.
• Outputs: native depth, supported disparity and validity conventions identified.
• Coverage: interior and thin-feature regions evaluated separately.
• Stability: valid-frame availability and depth spread reported independently.
• Motion: lag, disappearance and stale-value behavior checked.
• Accuracy: measurements compared with a suitable known reference.
• Processing: inferred or smoothed values distinguished where supported.

Keep the least aggressive processing that meets the task's measured acceptance limits. If no configuration preserves the required feature, reconsider viewing geometry or sensing method rather than hiding the failure with filling.

Start a camera configuration once those requirements are defined, or send the captures and checklist through application support for help narrowing the evaluation.