Planar tag pose ambiguity: diagnose flickering estimates

A stationary planar tag can produce a steady image while its estimated orientation jumps between two plausible poses. The camera position expressed in tag coordinates may then appear to cross an axis, even though nothing moved. That symptom does not, by itself, prove a faulty homography calculation.

This troubleshooting workflow separates coordinate mistakes, corner errors and image degradation from genuine geometric ambiguity. You will learn how to compare candidate poses, test physical constraints and reject uncertain measurements without hiding the problem behind a translation-sign filter.

1. Identify what actually changes

Start by recording the detected corners and pose before applying filters. Save the image, timestamp, camera calibration, target dimensions and coordinate-frame definition with each measurement.

Distinguish three different observations:

• The displayed axes flip: the pose may be changing, or the visualization may use an incorrect transform.
• The camera position changes sign: inversion can amplify an orientation change into a lateral position change.
• The projected corners move away from the detection: suspect correspondence, calibration or solver failure before accepting geometric ambiguity.

Use the existing pose estimation from detected corners workflow to check the basic input pipeline. The investigation here begins when that pipeline produces intermittent, competing answers.

Fix deterministic input mistakes before tuning temporal filters. Otherwise, filtering can make a wrong pose look convincing.

2. Declare transform direction before testing signs

Write down the exact equation your solver returns. A common convention maps target coordinates into camera coordinates:

Here, R_CT and t_CT describe the target-to-camera transform. C_T is the camera origin expressed in target coordinates. It is not the same quantity as t_CT, and inversion is not simply negating translation.

Check your implementation against the reference-frame explanation for pose initialization. Also confirm matrix storage, multiplication order and whether coordinates are column vectors or row vectors.

Record these conventions explicitly:

• Which direction each camera axis points.
• Which target corner defines the origin.
• Which directions define the target axes and printed face.
• Whether the reported transform maps target to camera or camera to target.
• Which length unit the object points use.

Under a camera convention with positive depth pointing forward, every visible target corner must have positive camera-frame depth. Test all corners, not only the target origin. If your camera convention differs, adapt the visibility test accordingly.

A sign constraint on the camera position in target coordinates is a different test. It is valid only when that axis and the permitted operating region are independently established. A camera being physically below a tag does not automatically imply a particular translation sign in an undocumented frame.

3. Eliminate corner and calibration mismatches

Overlay the corner index beside every detected corner. The object-point sequence must match the detector's documented corner identities, including the tag's decoded orientation.

Do not independently sort image corners by their current screen positions. A rotated tag can change which corner appears uppermost without changing its physical identity. Likewise, reversing the winding of only one point array creates an inconsistent correspondence.

Next, check that the calibration belongs to the image supplied to the solver:

1. Confirm the acquisition resolution and any crop origin.
2. Adjust intrinsics correctly when resizing or cropping.
3. Pair raw images with the appropriate distortion model.
4. Pair rectified images with their corresponding intrinsics and distortion treatment.
5. Confirm the physical target dimension matches the solver's object-point definition.

The rectified-image intrinsics guide addresses a common source of otherwise plausible errors. For persistent scale errors, use the distance and calibration checks.

Reproject the same object corners through every candidate using one consistent camera model. For a square target, define a corner reprojection score explicitly:

This definition is the root mean square of the corner-distance residuals. A library may report a differently normalized quantity, so do not compare error values until their definitions match.

Also inspect individual residual vectors. A low average can conceal one damaged corner, while a consistent directional pattern can reveal a calibration mismatch.

4. Test whether the ambiguity is genuinely geometric

A planar target supplies strong information about its image location and scale, but some viewing geometries provide weak information about its tilt. Near-frontal views and small image footprints can make two different orientations produce very similar corner projections.

Finite perspective usually distinguishes the alternatives in ideal data. In practice, corner noise can exceed that distinction. Two candidate poses may therefore have positive corner depths and similarly small reprojection errors.

Use a planar-pose method that exposes competing candidates where the documented interface supports it. A single returned solution does not prove that no other plausible solution exists.

Run a controlled stationary test:

1. Capture 100 frames without moving the target or camera.
2. Save the detected corner coordinates for every frame.
3. Evaluate candidate poses using the same calibration and error definition.
4. Plot reprojection error, target-normal direction and camera position separately.
5. Repeat after changing the target angle or increasing its image footprint.

If switching decreases when the view contains stronger perspective cues, geometric ambiguity becomes more likely. If identical corners produce inconsistent answers on repeated calls, investigate initialization, numerical settings, shared memory and solver state instead.

Do not interpret a small reprojection advantage as strong certainty. An error of 0.18 px versus 0.20 px may be insignificant relative to the actual corner uncertainty.

5. Quantify the available image evidence

Consider an illustrative square target with a physical width of 40 mm. Suppose its projected width is approximately 160 px and the calibrated focal length is 1000 px. For a nearly frontal target, the approximate distance is:

A width measurement change of 1 px gives an approximate relative distance change of 1 / 160 = 0.00625, or 0.625%. At the estimated distance, that is approximately 250 × 0.00625 = 1.6 mm.

This is a scale-sensitivity estimate, not a complete pose uncertainty model. Tilt can remain poorly constrained even when apparent width and estimated distance look stable.

You can also calculate the required image width from the target footprint. Suppose the horizontal field of view is 240 mm, and you choose 160 px across the target as an initial trial objective:

The chosen footprint is an application assumption, not a universal acceptance threshold. Validate it against your required angular accuracy and actual corner quality.

Increasing pixel count helps only when the optics, focus and contrast preserve useful edge information. Enlarging a blurred image after capture does not add independent geometric evidence.

6. Rank candidates with geometry and motion together

Apply checks in a defined sequence rather than accepting the first pose that passes a sign test:

1. Reject invalid numerical outputs and invalid rotations.
2. Reject candidates that violate the declared visibility convention.
3. Reject candidates outside independently known workspace or orientation bounds.
4. Compare reprojection residuals against measured image uncertainty.
5. Compare remaining candidates against feasible relative motion.
6. Report uncertainty if the evidence still does not separate them.

For temporal checks, compare poses in the same reference frame. Do not compare target-to-camera translation in one frame with camera-in-target position in the next.

Use a rotation-aware angular difference, not subtraction of Euler angles. Euler-angle representations can jump at wrapping boundaries even when physical orientation changes smoothly. Similarly, do not average rotation-matrix elements without restoring a valid rotation.

As an illustrative motion gate, suppose the maximum relative translation speed is 0.2 m/s and the elapsed time is 0.02 s:

A relative angular speed bound of 60 degrees/s over the same interval permits 60 × 0.02 = 1.2 degrees. Add allowances for measurement uncertainty, and derive the limits from actual machine motion rather than these example values.

Use measured timestamps, especially after dropped frames. Longer gaps allow larger legitimate changes. A temporal gate also needs a trustworthy initial state: choosing the wrong branch at startup and preserving it indefinitely is not validation.

Ordinary tag tracking is not a safety-rated protective function. Keep machine safety independent of this pose-validation logic.

7. Improve acquisition before replacing the solver

Repeat the test with fixed exposure and gain, stable illumination, accurate focus and a clean target. Inspect the corners at native image resolution rather than judging only the full-frame preview.

Motion blur weakens edge localization. Rolling-shutter readout can also distort a moving target because different image rows represent different times. Neither effect is equivalent to ordinary corner noise, and a single rigid-pose model may no longer describe the observed image well.

Use the exposure, gain and balance setup guide when acquisition settings are changing between frames. Then browse area scan cameras and the camera finder only if the remaining limitation is acquisition-related.

For a global-shutter acquisition option, review the MV-CU020-90GC. That shutter type can address motion-related distortion; it does not remove planar pose ambiguity.

The LS-CMR-0166 is a rolling-shutter alternative to evaluate when the target remains stationary during acquisition. Confirm the complete timing behavior for your application before choosing it.

These are acquisition candidates, not a claim of pose accuracy. Target footprint, lens performance, calibration and geometry still determine whether the image supports the required measurement.

If you need independent depth evidence, assess 3D sensors. Their suitability depends on measurement range, surface behavior, uncertainty and alignment with the tag image; a depth value alone does not establish the target's full orientation.

8. Verify the fix with a repeatable acceptance test

Build a test set that contains the difficult geometry, not just favorable examples. Include stationary views, near-frontal views, tilted views and the fastest permitted motion.

For each condition, retain the raw measurements and check:

• Corner identity: labels stay attached to the same physical corners.
• Projection quality: residuals remain consistent with the image evidence.
• Visibility: all corners satisfy the declared camera-depth convention.
• Pose continuity: changes remain within independently justified motion bounds.
• Ambiguity reporting: unresolved candidates produce an uncertainty state.
• Recovery: tracking reacquires correctly after occlusion or a timestamp gap.
• Accuracy: estimates agree with an independent reference where available.

Track false acceptance separately from rejected measurements. A filter that suppresses visible jumps by rejecting most frames may be unsuitable for a control loop requiring frequent updates.

For example, examine 1000 frames and count accepted, rejected and independently identified wrong poses. Report the operating conditions with those counts; a single overall percentage can conceal failure at one critical target angle.

9. Know when a different target arrangement is needed

If the required angular accuracy cannot be supported near a frontal view, software tuning may not be the right remedy. Increase the target footprint, improve viewing geometry, or introduce known non-coplanar features.

Multiple tags on the same flat board improve spatial support, but they remain planar. They do not automatically provide the depth diversity of a rigid, non-coplanar arrangement. Likewise, temporal smoothing cannot recover information absent from a stationary image.

Before ordering, collect the target dimensions, field of view, working distance, motion limits and required pose uncertainty. Use the camera configuration step for acquisition selection, or contact applications support with representative images and candidate-pose logs. The next decision should follow the failed test: correct the data pipeline, strengthen the geometry, or change acquisition hardware.