Choose stereo camera baseline distance for 3D vision

Two separate cameras can form a stereo pair, but there is no universal spacing that makes them accurate. A useful baseline depends on your working range, image scale, required depth uncertainty and how much of the part both cameras must see. This guide shows how to calculate a starting distance, compare practical alternatives and validate the assembled pair before fixing the mounts. The calculations concern measurement, not comfortable viewing on a stereoscopic display.

Start with depth uncertainty and shared coverage

The baseline is the distance between the cameras’ optical centers. It is not necessarily the gap between their housings or the distance between their mounting holes.

Increasing baseline generally improves depth sensitivity: a given change in object distance produces a larger disparity change. However, larger separation also increases viewpoint differences, hidden surfaces and the image displacement your matching software must search.

Write down these requirements before choosing a dimension:

• Nearest and farthest working distances, measured from a defined camera coordinate origin.
• Required common inspection width at each distance.
• Smallest feature that must remain recognizable in both images.
• Acceptable random depth uncertainty and allowable systematic depth error, recorded separately.
• Whether the object, cameras or lighting change during acquisition.
• Available mounting space and adjustment access.

Do not copy human eye spacing as a measurement rule. A baseline suitable for a nearby textured object may be inadequate for a distant surface or excessive around deep recesses.

If your main decision is which camera architecture suits a large separation, use the related wide-baseline stereo camera selection guide. Here, the starting point is the required geometry rather than a camera format.

Calculate baseline using calibrated image geometry

For a rectified stereo pair with horizontal disparity, the basic relationship is:

Here, Z is depth along the rectified camera axis, B is baseline, and f_px is focal length expressed in pixels. Disparity d is the horizontal image displacement after accounting for any difference between the rectified principal points. Use the same length unit for depth and baseline.

The uncertainty equation is a first-order estimate. It assumes a small disparity error, fixed calibration and a valid correspondence between the same surface point in both images. It does not include calibration bias, mechanical movement or incorrect matches.

For preliminary planning with parallel cameras, estimate image scale from horizontal field width:

N_x is the processed image width in pixels, and W_ref is its horizontal field width at reference depth Z_ref. Use dimensions from the same image coordinate system. Resizing, rectification and cropping can change the relationship between pixel coordinates and field coverage.

Obtain the final focal value from calibration rather than treating a lens label as sufficient. For an existing installation, calculating angle of view from calibration helps connect the calibrated model to physical coverage.

The disparity uncertainty, sigma_d, must also be justified. A provisional subpixel value is a planning assumption, not a matcher specification or an accuracy guarantee. Measure it on representative surfaces before approving the design.

Work through a baseline example

Consider a stationary inspection with these application assumptions:

• Processed image width: 4000 pixels.
• Horizontal field width: 600 mm at a reference depth of 1000 mm.
• Working depth range: 800–1200 mm.
• Required common inspection width: 350 mm.
• Target random depth uncertainty: 0.5 mm at the farthest distance.
• Provisional disparity uncertainty: 0.2 pixel.

First estimate the focal length in pixel units:

Then calculate the baseline needed at the farthest working distance:

A trial baseline of 90 mm provides a convenient starting point. Its predicted random depth uncertainty at the far distance is:

The same assumptions predict about 0.21 mm at the nearest distance. This difference follows the squared dependence on depth; it is not evidence that the assembled system will achieve either result.

Now check sampling separately. At the far distance, the estimated field width is:

The assumed image width clears that sampling target, although contrast, focus and matching texture still need verification. Use the camera finder to screen image dimensions, then check the actual processed output rather than relying on the camera’s headline resolution.

Compare baseline against overlap and disparity range

For identical, parallel cameras with equal horizontal coverage, a first estimate of common width on a frontoparallel plane is:

This assumes aligned optical axes and neglects rectification cropping, lens distortion and occlusion. It describes geometric overlap on a plane, not the percentage of pixels that will produce valid depth.

In the worked example, each camera covers approximately 480 mm at the nearest plane. The following comparison uses the same image scale and disparity uncertainty throughout.

The middle option leaves 40 mm of total geometric margin around the required inspection width at the nearest plane. Rectification cropping and alignment errors consume some of that allowance. The larger option leaves only 10 mm before those losses.

Next, calculate the expected disparities for the middle option:

The matcher must support this disparity interval, with additional allowance for the real geometry and calibration. Check both its search-range width and any supported minimum-disparity offset. Large displacement can increase processing cost even when the images contain enough texture.

For coverage losses after processing, the guide to why depth-image field of view is smaller explains why two complete camera images do not imply an equally complete depth image.

Choose cameras and mounting hardware as a system

Start with area scan cameras whose image dimensions, shutter behavior and acquisition controls fit the task. Matched camera and lens configurations usually simplify integration, but identical labels do not prove synchronized exposure or compatible software operation.

The LS-CMR-0166 is a USB3 area scan camera with 4000 × 3000 resolution and a rolling shutter.

The LS-CMR-0053 is a 12 MP GigE area scan camera with a rolling shutter.

These are individual camera candidates, not a validated stereo kit. Treat them as alternatives, not a recommendation to mix them into one pair. Confirm manufacturer-documented triggering, exposure timing, image access and simultaneous operation before selecting either configuration.

Rolling-shutter cameras require particular care when the scene moves. Frame timestamps alone do not establish simultaneous exposure of corresponding rows. For moving parts, consider global-shutter candidates and verify the complete trigger and exposure timing chain.

Use camera stands and a common rigid support to establish repeatable geometry. Independent stands can be useful during experiments, but relative movement between them changes the stereo calibration.

The LS-MNT-0040 is a professional slide-table camera stand.

A stand listing alone does not establish suitability for a stereo assembly. Confirm the mounting interfaces, supported load, adjustment locking and provision for the second camera. The accessories finder helps identify mounting components, but mechanical suitability still requires checking the assembled structure.

Provide enough adjustment to establish the trial baseline and aim both views. Then lock the assembly before calibration. Do not assume an adjustment scale measures the optical-center separation accurately.

Calibrate the fixed pair and verify independent distances

Calibration determines the actual relative camera geometry; it should not merely confirm the nominal spacing on a drawing. Follow the paired-image stereo calibration procedure after fixing focus, image dimensions and mounting positions.

Use this validation sequence:

1. Calibrate each camera and the pair. Capture a target across the usable image area, with varied orientations and distances. Preserve the exact image mode used for inspection.
2. Check rectification. Corresponding features should align vertically in the rectified images. Inspect the center and edges, not just one convenient patch.
3. Measure a representative textured target repeatedly. Keep it stationary and record valid depth, spread and invalid-pixel rate at the near, middle and far positions.
4. Compare against an independent distance reference. Account for the offset between the reference datum and the rectified camera origin. Keep axial depth distinct from slant range.
5. Repeat after normal operation. Check whether warm-up, adjustment locking or routine handling changes the results.

Choose a target that covers the relevant image area without clipping. The calibration target coverage and margin guide helps plan that geometry. Target shape matters too: use the target flatness guide when deciding whether a printed sheet is sufficient.

For repeated stationary measurements, calculate depth bias and spread separately:

Compare these results against separately defined acceptance limits. A low spread with a large bias indicates a different problem from unstable matching. Likewise, a low calibration residual does not replace an independent accuracy test; see extrinsic calibration accuracy beyond reprojection error.

Know when a wider baseline is the wrong fix

A wider baseline cannot recover a surface visible to only one camera. Deep pockets, overlapping parts and sharp foreground edges can lose common visibility as separation increases. Test representative geometry, not only a flat calibration board.

Depth boundaries also violate the assumptions behind simple uncertainty estimates. A matching window can combine foreground and background texture, producing displaced or unstable edges. The related stereo depth edge accuracy guide explains why edge errors need their own acceptance checks.

Low texture, reflections and repeating patterns can produce false matches even when the calculated baseline looks suitable. Improve the image formation and matching conditions before attributing every error to camera spacing. More separation can make the views less similar and worsen correspondence.

If you toe the cameras inward to increase central overlap, recalibrate the complete arrangement. Toe-in changes distortion, rectification and coverage tradeoffs; the simple parallel-camera overlap estimate no longer applies directly.

Finally, separate software problems from mechanical ones. A wrong rectification map, changed image scale or mismatched frame pair can resemble a poor baseline. If the common image area cannot contain a calibration target reliably, fix capture and framing before increasing separation.

Check the design before committing to hardware

Use this final checklist to turn the calculation into an orderable, testable arrangement:

• Working range defined: use the far distance for uncertainty and check near-plane overlap separately.
• Image sampling checked: verify the smallest required feature at the largest field width.
• Disparity assumption measured: replace the planning value with representative test results.
• Matcher limits confirmed: check disparity offset, search width, processing time and invalid-depth behavior.
• Acquisition timing verified: confirm actual exposure behavior for both views, especially with motion.
• Mechanical geometry stable: lock focus, camera spacing and aiming before final calibration.
• Coverage margin retained: allow for rectification cropping, mounting tolerances and the real part outline.
• Acceptance evidence recorded: retain independent distance checks, repeatability results and edge tests.

Choose the smallest baseline that meets the validated depth requirement while retaining sufficient shared visibility and mounting stability. Do not select the widest spacing simply because its ideal uncertainty calculation is better.

Start a candidate assembly in the camera configurator. If the calculation leaves little overlap or uncertainty margin, send your working range, required inspection width and measured error limits through applications support before fixing the mechanical layout.