Monocular reconstruction scale ambiguity: fixing large values

Large coordinates from monocular structure-from-motion can be harmless, but they can also hide an unstable reconstruction. Coordinate magnitude alone cannot distinguish those cases. This troubleshooting procedure helps you separate arbitrary scale and origin choices from tracking errors, weak triangulation, and uncertain camera calibration. You will establish physical units with a known distance, transform camera poses consistently, and verify geometry using measurements that did not set the scale.

Start with the symptom, not the coordinate magnitude

A reconstruction expressed in arbitrary units may contain very large coordinates without representing a physically enormous scene. First inspect relative distances, camera positions, and reconstruction stability rather than individual coordinate components.

Work through these causes in order. The order is a practical diagnostic sequence, not a probability estimate for every dataset.

If shape remains stable and independent measurements agree after one uniform scale change, the original coordinate magnitude was not itself the problem. If different frame selections produce different shapes, scaling is not a sufficient repair.

Understand what monocular geometry cannot determine

For a calibrated perspective camera observing a rigid scene, monocular reconstruction generally recovers geometry only up to a global similarity transformation. Absolute position, orientation, and scale require external choices or information.

In camera-center notation, perspective projection can be written as:

Here, X is a scene point, C is the camera center, R maps world directions into camera coordinates, and K contains the intrinsics. The symbol ~ means equality up to homogeneous scale.

Multiplying both X and C by the same positive factor leaves their image projection unchanged. Translating both by the same vector also leaves their relative geometry unchanged.

Consequently, image tracks alone cannot tell you whether a reconstructed spacing represents millimeters or meters. Increasing image resolution or processing more frames does not supply that missing physical reference.

There is an additional qualification: with unknown intrinsics and insufficient self-calibration constraints, the ambiguity can be more general than a uniform scale. A single known length cannot repair an incorrectly shaped reconstruction. The discussion of singular calibration estimates explains why an optimizer can lack enough independent information.

Audit tracks and coordinate conventions

Before adjusting solver settings, inspect the actual observations supplied to reconstruction.

1. Overlay tracks on the source images. A track identifier must follow the same physical feature throughout its visible lifetime.
2. Check coordinate ordering. Confirm horizontal and vertical coordinates are not exchanged, and verify the image-origin convention.
3. Audit resized or cropped frames. Intrinsics and feature coordinates must describe the same image geometry. Cropping shifts the principal point; resizing changes focal lengths expressed in pixels.
4. Represent missing observations correctly. Use the documented missing-data representation. A fabricated observation at the image origin can corrupt geometry.
5. Confirm the scene is sufficiently rigid. Independently moving objects do not obey the same static-scene camera model.
6. Check the returned point representation. If an output is homogeneous, convert using its documented final coordinate. Do not apply homogeneous division to an already Euclidean point.

For example, halving image dimensions requires corresponding changes to focal lengths and principal-point coordinates expressed in pixels. Updating the image without updating its calibration creates an inconsistent projection model.

If frames were resampled, the guide to antialiasing during downsampling and image warping helps distinguish resampling artifacts from tracking problems.

Finally, confirm whether an exported pose contains a camera center or an extrinsic translation. They are not interchangeable. For the convention X_camera = R X_world + t, the camera center is C = -Rᵀt.

Test whether the footage contains useful parallax

Many frames do not necessarily provide many independent depth constraints. Consecutive images can differ mainly through rotation, or contain too little translation relative to scene distance.

Pure camera rotation does not provide triangulation depth. Forward motion can also constrain points near the direction of travel poorly, even when peripheral features move visibly.

For a simplified rectified lateral pair, depth sensitivity is approximately:

B is the baseline, d_px is disparity, and the uncertainty expression considers disparity error only. It is an illustration of triangulation sensitivity, not a complete uncertainty model for an arbitrary camera trajectory.

With disparity of 2 pixels and disparity uncertainty of 0.2 pixels, relative depth uncertainty is approximately 0.2 / 2 = 10%. At 20 pixels, the same calculation gives approximately 1%.

For your actual sequence:

• Inspect viewing-ray angles for reconstructed points.
• Compare reconstructions using several overlapping frame subsets.
• Select frames with useful translational separation while retaining reliable feature correspondence.
• Reject points whose depth changes dramatically when a few observations are removed.

There is no universal triangulation-angle threshold that guarantees measurement accuracy. Choose acceptance limits from your required uncertainty and verify them experimentally. A distant feature may track reliably while remaining poorly constrained in depth.

Separate calibration uncertainty from scale ambiguity

An assumed field of view supplies an initial guess, not a camera calibration. Intrinsic refinement can improve that guess only when the observations and motion contain enough information.

Allowing focal length, principal point, and distortion to vary simultaneously can introduce strong parameter coupling. A solver may explain the observations using a plausible-looking but incorrect combination of camera parameters and scene geometry.

When you control acquisition, calibrate at the intended focus, image size, and lens settings. Keep those settings fixed during reconstruction. The guide to calibration-target coverage margins explains why observations should cover the useful image area rather than cluster near the center.

For existing footage without camera information, compare solutions under several defensible intrinsic assumptions. Large changes in recovered shape indicate uncertainty that one scale reference cannot remove. Digital stabilization, changing zoom, or rolling-shutter motion can also violate a simple fixed-intrinsic, single-pose-per-frame model.

A tiny reprojection residual is not a substitute for these checks. The closest measurement-focused discussion, camera calibration accuracy beyond reprojection error, explains why image-space agreement and physical accuracy are different tests.

Recover metric scale from a known scene distance

Once geometry is stable, choose a physical distance between identifiable scene points. Measure the same endpoints in the reconstruction. Prefer a long, well-defined reference with reliable tracks and well-constrained depth.

Let L be the measured physical length and d_raw its reconstructed length:

The vector o is an optional origin chosen in the raw reconstruction frame. These equations change scale and origin without rotating the axes.

Suppose a reference measures 1,200 mm physically and spans 2.4 × 10^8 reconstruction units:

A second reconstructed spacing of 3.6 × 10^7 units then becomes:

Apply the same transformation to every scene point and camera center. Under the stated convention, rotations remain unchanged and extrinsic translations must be recomputed:

Scaling scene points while leaving camera positions unchanged breaks projection consistency. If you also rotate the world axes, update camera rotations accordingly.

Use several reference distances when practical. Compare their individual scale estimates before combining them. Disagreement beyond measurement uncertainty suggests geometric distortion, endpoint mistakes, or inadequate triangulation—not a need for separate scale factors along different axes.

Verify accuracy independently of the scale reference

The distance used to establish scale will agree by construction. Reserve other physical measurements for validation, preferably at different depths and positions across the scene.

Suppose an independent distance measures 600 mm, but its raw reconstructed length is 1.16 × 10^8 units. Using the previous scale:

The reference distance can be exact while this independent measurement remains unacceptable.

Use three complementary checks:

• Physical checks: compare withheld lengths and their uncertainty against the application tolerance.
• Image checks: independently project reconstructed points using the documented camera model and compare with the observed image coordinates.
• Stability checks: repeat with different frame subsets, initializations, and reasonable intrinsic constraints.

For image checks, confirm that observations and predictions use the same pixel coordinate system and distortion convention. Recompute residuals rather than assuming a solver log uses your expected definition or units.

Also check positive depth in observing cameras and inspect residuals by frame and image location. One aggregate error can hide systematic patterns. The guide to length measurement from image markers provides a complementary approach when your actual task is measurement on a known plane.

Improve the next capture before changing hardware

Capture a static, textured scene with overlapping views and deliberate translational motion. Keep focus and image processing fixed. Include a measured reference and separate validation dimensions whenever possible.

Reference visibility also needs adequate sampling. For an illustrative planar view spanning 800 mm, a target sampling of 0.2 mm per pixel requires:

This sizes image sampling only. It does not predict depth accuracy or replace the parallax checks.

Use the guide to resolution for dimensional inspection to relate sampling to your smallest relevant feature. Compare area scan cameras in the camera finder once that requirement is clear.

The LS-CMR-0053 is a rolling-shutter camera option to evaluate for controlled captures with the camera and scene stationary during exposure. Confirm acquisition behavior and software compatibility before selecting it.

Review machine vision lenses through the lens finder, checking coverage, working distance, mount, and focus stability. The LS-FFL-0567 is a C-mount lens option to evaluate against those requirements.

Neither component establishes metric scale by itself. A documented capture model, observable geometry, and physical reference remain necessary.

Close the fault with a documented acceptance check

Before accepting the reconstruction, record the following:

• Track identity, missing-data handling, and image transformations have been checked.
• Camera intrinsics match the processed images, or their uncertainty is explicitly documented.
• Translation provides useful parallax for the points being measured.
• One uniform scale and a defined origin are applied consistently to points and camera poses.
• Independent dimensions pass the application tolerance.
• Geometry remains stable under reasonable reconstruction changes.
• Display normalization is kept separate from the measurement transform.

Do not accept a scaled reconstruction when the underlying scene deforms, the camera motion is effectively rotational, or plausible calibration changes produce incompatible shapes. In those cases, acquire better observations or use a measurement method with additional physical constraints.

For a repeatable acquisition setup, start with the camera configurator. If the footage or geometry remains uncertain, bring representative frames, track overlays, and reference measurements to application support before selecting replacement hardware.