An object can produce a clean depth outline nearby and a ragged, sloping boundary farther away, even with unchanged camera settings. For dimensional inspection, the question is not how to make that outline look smoother. You need to establish whether the boundary is measured accurately, missing, or reconstructed incorrectly. This troubleshooting procedure separates stereo geometry, boundary occlusion, processing artifacts, and display effects. It also shows when fitting supported surfaces is more defensible than measuring individual edge pixels.
Start with the symptom, not a smoothing filter
Stereo depth edge accuracy depends on correspondence between two camera views. At an object boundary, those views may not contain the same surface. Increasing distance also reduces spatial sampling and increases the depth error associated with a given disparity error.
Several mechanisms can therefore appear together. Diagnose them separately before changing settings.
| Observed symptom | Likely cause | First test |
|---|---|---|
| Flat surfaces become noisier with distance | Disparity uncertainty amplified by range | Compare repeated measurements of an interior surface patch |
| A narrow band beside the object contains no valid depth | Occlusion or rejected stereo matches | Inspect both rectified source images and the validity mask |
| Points bridge the object and background | Incorrect matches or processing across the discontinuity | Compare unfiltered depth with processed output |
| Boundaries change after color registration | Resampling, visibility differences, or calibration errors | Inspect the native depth grid before alignment |
| Only the rendered outline changes | Point size, triangulation, color scaling, or perspective | Inspect numeric coordinates without mesh reconstruction |
For missing samples, use the separate procedure for missing points in a 3D camera point cloud. For boundaries that change specifically during registration, follow lines in aligned depth images.
A rejected depth sample is not necessarily a defective measurement. At an occluded boundary, withholding a result can be more accurate than returning a plausible-looking coordinate.
Calculate what doubling the distance changes
For rectified stereo geometry, axial depth follows disparity. A first-order uncertainty estimate explains why the same matching uncertainty has a larger effect farther away.
Here, Z is axial depth, f_px is rectified focal length in pixels, B is stereo baseline, and d is disparity in pixels. Keep depth and baseline in the same length unit. The uncertainty estimate assumes small disparity errors and valid correspondence; it does not describe gross mismatches at boundaries.
Consider hypothetical geometry with a focal length of 700 pixels, a baseline of 50 mm, and disparity uncertainty of 0.2 pixels. These are illustrative values, not specifications for a listed product.
At a depth of 350 mm:
At a depth of 700 mm:
Doubling distance produces four times the estimated depth uncertainty if disparity uncertainty remains unchanged. It also doubles the object-space width represented by each pixel. Actual performance may deteriorate differently because texture, exposure, and matching confidence also change.
| Quantity in this hypothetical example | Near position | Far position |
|---|---|---|
| Axial depth | 350 mm | 700 mm |
| Disparity | 100 pixels | 50 pixels |
| Estimated depth uncertainty | 0.7 mm | 2.8 mm |
| Approximate object-space sampling | 0.5 mm/pixel | 1.0 mm/pixel |
A smoother display cannot recover the spatial information lost through reduced sampling. Likewise, improved texture may reduce matching uncertainty without eliminating the underlying range dependence.
Run a controlled straight-edge test
Use a rigid target with a known straight boundary and independently measured dimensions. A matte rectangular block on a flat reference surface is useful, provided its edge straightness is better than your inspection requirement. Avoid flexible packaging or rounded corners when establishing the baseline.
- Define the measurand. Decide whether you need top-face width, edge straightness, height above the reference surface, or a complete side profile. These are different measurements.
- Establish reference geometry. Locate the expected edge independently in the measurement coordinate system. Fitting a line to the measured points alone cannot reveal a uniformly shifted edge.
- Record the acquisition configuration. Save stream dimensions, calibration, depth scale, exposure, confidence settings, projector state where applicable, and processing settings.
- Capture native data. Save the stereo source images where available, native depth, validity information, and timestamps. Initially disable optional smoothing, hole filling, and temporal accumulation using documented controls.
- Test multiple distances. For example, acquire at 350 mm, 500 mm, and 700 mm while maintaining comparable target orientation and illumination.
- Repeat each acquisition. A practical starting dataset is 100 stationary frames per distance. Keep the target and fixture stationary throughout each sequence.
- Measure the same physical regions. Use an interior surface patch, a boundary band, and a background patch. Do not compare unrelated pixel rectangles as the target changes apparent size.
Before comparing datasets, follow the checks in raw stereo image calibration. Calibration and stream metadata must correspond to the images actually processed.
Record signed depth bias on the interior patch, temporal variation, valid-sample percentage near the boundary, and lateral edge displacement. Report a percentile such as the 95th-percentile absolute error alongside coverage, rather than hiding rejected samples inside an average.
Separate measurement errors from display effects
Begin with numeric depth values and a validity mask, not a shaded point-cloud screenshot. Fix the depth display range across test distances so automatic color scaling does not exaggerate or conceal changes.
Inspect the cloud as unconnected points before generating a mesh. A mesher can join foreground and background samples across missing data, creating a sloping curtain that the camera never measured. Larger rendered points can also close visual gaps without adding information.
Use equal axis scales and a consistent viewing direction. Then inspect numeric cross-sections perpendicular to the known edge. Plot lateral position against axial depth, retaining invalid samples as gaps rather than replacing them with zeros.
Compare the native cloud with any cloud exported by the application. An export may already contain filtering, decimation, registration, or triangulation. Seeing the same shape in another viewer does not prove that the underlying acquisition was unprocessed.
Changing depth encoding resolution is another common distraction. Encoding depth in increments of 0.1 mm rather than 1 mm can reduce storage quantization where supported. It does not make the optical measurement ten times more accurate. Read the documented scale from metadata and check the representable range and invalid-value convention.
If units or clipping appear inconsistent, use depth scale mismatch diagnosis. Do not alter scale merely to make the object appear dimensionally correct.
Fix acquisition and matching before filtering
Work through the likely causes in this order, changing one variable per comparison.
First, check exposure and usable texture. Inspect the actual stereo images, not just the color image. Saturated highlights, dark regions, repetitive patterns, and nearly featureless surfaces can all weaken correspondence. When supported, compare documented projector settings while staying within the manufacturer's operating limits. Opening an image stream does not itself establish that illumination is active.
Second, check boundary visibility. A foreground object hides different background regions from the two viewpoints. A background patch visible to only one camera cannot provide a normal stereo correspondence. A larger baseline can improve depth sensitivity while increasing some occlusion regions; it is not a universal boundary fix.
Third, inspect matching confidence and rejection. Less restrictive settings may return more points but admit more outliers. More restrictive settings may improve accepted-point quality while leaving larger gaps. Compare both error and coverage, using the procedure for stereo disparity map troubleshooting when correspondence itself looks unstable.
Fourth, restore processing selectively. Evaluate each filter against the reference dimensions. Temporal filtering can stabilize stationary surfaces but leave motion trails. Spatial filtering can reduce noise but move discontinuities or mix surfaces. Hole filling estimates missing geometry; it does not provide an independently observed edge.
If the target moves, repeat validation at production speed. A configuration that passes a stationary test may fail because of exposure blur, synchronization differences, or temporal processing.
Use fitted geometry only when the assumptions are valid
For a rigid prismatic part, the physical edge may be better estimated from supported surfaces than from isolated boundary pixels. Fit planes to reliable interior points away from the discontinuity, then calculate their intersection when both surfaces are sufficiently observed.
This produces a model-based estimate, not a direct measurement of every edge sample. Validate it against independently measured parts across the intended pose and distance range. Report fitting residuals and geometric conditioning; nearly parallel planes produce poorly constrained intersections.
A top view may reveal the top face but little or none of a vertical side. You cannot claim a measured top-to-side intersection if the side plane was assumed rather than observed. A nominal box model may still be useful, but its dimensional assumptions must remain explicit.
Another option is to detect an outline in a calibrated image and intersect its viewing rays with a validated object plane. For an elevated top face, that plane must represent the top face, not the supporting table. Using the table plane can introduce systematic dimensional errors.
The geometry behind ground-plane pose from known points helps establish a reference frame, but it does not automatically locate elevated edges. Validate the complete transformation using the checks in camera extrinsic calibration accuracy.
Use raw boundary measurements instead when chamfers, damage, curvature, or deformation are themselves inspection targets. An idealized plane model can conceal exactly those defects.
Choose camera geometry around the required tolerance
If the working distance is fixed, evaluate focal length in pixels, baseline, usable field of view, and accepted-point coverage together. More output pixels help only when the optical image and correspondence process preserve the necessary detail. Interpolated output does not create additional measurements.
For a separate hypothetical sampling calculation, suppose the field width is 640 mm and the desired sampling is 0.5 mm/pixel:
This is a sampling requirement, not an accuracy guarantee. A minimum feature should normally span multiple samples, and the required margin depends on contrast, localization method, and tolerance.
The relationship between field of view, working distance, and image coverage explains why moving the same camera farther away reduces sampling density.
Use the 3D camera working distance and positioning guide to define the geometry before comparing 3D sensors in the camera finder.
The MV-DLS300P-04 is identified in the catalog as a laser galvanometer stereo camera. Treat it as a candidate for application review, not as evidence of a particular boundary tolerance.
- Type
- Stereo / structured light
The MV-DPS200P-02 is identified as a projected structured-light camera. Its suitability for the required geometry needs confirmation through application data and representative samples.
- Type
- Stereo / structured light
Neither product description establishes comparative edge accuracy. Request evidence at your working distance, field of view, material, and measurement tolerance. If stereo cannot meet the requirement, evaluate another measurement geometry rather than assuming more filtering will solve it.
Set acceptance criteria before ordering
Use a short acceptance checklist that distinguishes appearance from metrology:
- Reference: Confirm target dimensions and straightness with a suitable independent measurement method.
- Range: Test the nearest, nominal, and farthest operating positions, including permitted part tilt.
- Coverage: Specify the minimum valid-sample percentage and the largest permissible gap near the boundary.
- Accuracy: Set separate limits for signed dimensional bias, temporal variation, and lateral edge error.
- Processing: Record every enabled filter and retain an unfiltered dataset for comparison.
- Materials: Include representative surface finish, color, and reflectivity rather than only an easy calibration target.
- Motion: Validate at the intended production speed when the application is not stationary.
- Model assumptions: Label plane-intersection or image-to-plane measurements as inferred geometry and validate them independently.
The usual answer—adding texture and smoothing—is wrong when occlusion or insufficient sampling dominates the error. It is also wrong when a fitted ideal shape would conceal defects that the inspection must detect.
Start a camera configuration with your required distance, field of view, and tolerance. For an application review, send representative raw captures and reference measurements through technical support so the next decision is based on measurement evidence rather than a cleaner-looking cloud.
Questions
Why does stereo depth edge accuracy get worse farther away?
For a fixed stereo geometry, a given disparity error produces greater axial depth error as distance increases. Each pixel also covers more of the object. Boundary occlusion and weak correspondence add further errors, so a nearby clean outline does not establish accuracy at a longer working distance.
Can hole filling make an edge accurate enough to measure?
Hole filling can make a cloud more complete, but it estimates samples that were not accepted as measurements. Near a foreground boundary, those estimates may extend the wrong surface. Use it for dimensional inspection only after independent tests demonstrate acceptable error and coverage across the intended operating conditions.
Does a smaller depth unit improve physical accuracy?
Not by itself. A finer encoding increment can reduce numerical quantization, but optical geometry, correspondence uncertainty, calibration, and surface visibility still determine measurement quality. Always interpret stored values using the documented depth scale, and check whether changing the encoding also affects the available depth range.
When should an edge be estimated from fitted planes?
Plane intersections are useful when the part genuinely has planar faces and enough reliable interior samples support each fit. They are less suitable when the inspection must detect rounded edges, dents, chamfers, or deformation. Treat the intersection as inferred geometry and validate its position against an independent reference.
What evidence should I request before choosing a replacement camera?
Provide the working-distance range, required field of view, part dimensions, surface properties, motion conditions, and dimensional tolerance. Ask for representative raw data and reference comparisons that report both accepted-point error and missing-data coverage. A visually smooth demonstration is not sufficient evidence of dimensional performance.