Skip to content

Degeneracy

Use this page when a calibration value looks wrong or varies between recordings, and you suspect the motion.

A motion is degenerate for a value when it gives the data no way to measure that value (Observability explains the reasons). A degenerate value does not make the solve fail. It converges, the residuals look fine, and the value sits near its start or drifts with noise. So you have to look for it.

Symbols follow Notation & glossary.

MotionWhat is lostHow it shows
Pure translation (no rotation)lever arm (all 3 axes); gyro scale Dw\mathbf{D}_w; the extrinsic rotation is fixed only by the direction of gravity and acceleration, so it is weak; readout and time offset are weaklever stays near its start with a wide 1σ; Dw\mathbf{D}_w diagonal stays near 1
Rotation about one axis onlythe lever-arm component along that axis; extrinsic rotation about that axis; Dw\mathbf{D}_w columns of the other two axes; readout, if the axis is the optical axisone lever component with a much wider 1σ than the others; the rotation-excitation check shows one large singular value and two tiny ones
Constant velocity (steady turning and moving)time offset td,it_{d,i} (a time shift looks like a fixed transform); t_ga (no change in acceleration); Da\mathbf{D}_a gets no signal beyond the gravity tiltstd,it_{d,i} and the lever move together between recordings
Planar motion (rig slides in a plane and turns only about its normal, as on a cart or a table)everything one-axis rotation loses; with a vertical normal, gravity stays on one IMU axis, so the other Da\mathbf{D}_a columns and the accel bias along gravity trade offas one-axis rotation, plus Da\mathbf{D}_a diagonals with a wide 1σ
Too slowtime offset, readout and t_ga (all need fast change); weak Dw\mathbf{D}_wwide 1σ on the timing values; readout far from the sensor’s spec
Too shortfewer distinct orientations; biases and IMU scales separate poorlyvalues move a lot between recordings
Board always centred, face-onlens distortion and focal lengthdistortion differs between recordings; residuals grow toward the image edges

Tg\mathbf{T}_g is weak under almost any hand-held motion. That is why it is pinned unless you set estimate.imu.intrinsics: full.

MVIS does not stop a run because the motion was weak. It reports, at four points.

  • Recording length: mvis inspect warns when a recording with an IMU is shorter than 30 s.
  • Window choice: for bags longer than 80 s, the camera–IMU run keeps the 80 s with the most gyro and accelerometer excitation, scored per IMU as the mean squared angular rate plus the variance of the acceleration magnitude (input.window_mode: motion).

The start-value step runs simple fits before the joint solve and checks them:

  • Rotation excitation: the singular values of the camera–gyro angular-velocity fit. One large and two tiny ones mean rotation about one axis.
  • Time-offset sharpness: how much the fit error rises away from the best offset. A flat minimum means the offset is weak.
  • Lever arm: if its 1σ in the linear fit is over 5 cm, it starts at 0.
  • Accelerometer bias: it starts at 0. The start step estimates it only when asked (an advanced start option); it then falls back to 0 when the linear fit cannot separate it (smallest singular value of the scaled system under 10⁻³, or 1σ over 0.05 m/s²).
  • Gravity and scale: the fitted gravity should be close to 9.81 m/s² and the scale within 3 % of 1. A miss points at wrong intrinsics, units or stamps.
  • Auxiliary IMUs: the gyro-to-gyro start needs a clear minimum (1.5 times sharper than the edges) and more than 10 s of overlap.

A value that starts at 0 is still solved jointly; the start only keeps a poor guess out.

Freed IMU intrinsics carry a weak prior (0.1 on Dw\mathbf{D}_w and Da\mathbf{D}_a). On a direction the motion does not excite, that prior wins: the value stays near its start instead of wandering. Its 1σ then stays close to the prior’s 0.1: that is the sign the motion did not excite it.

  • 1σ for every value: from the marginal covariance of the joint solve, in the result JSON and the report. Compare each with the others of its kind and across recordings.
  • Camera-only report, motion grades: each camera gets a grade (good, marginal, poor) per value, from its board poses. The thresholds are empirical:
Grade forMeasureGoodMarginal
focal lengthfarthest over nearest board distanceover 2.0over 1.5
distortionimage area covered by the board cornersover 60 %over 30 %
rotation diversitysmallest over largest spread of the rotation axesover 0.1over 0.03
time offsetspread of the angular speedover 0.5 rad/sover 0.2 rad/s
readoutpeak angular speed about the camera’s x and y axesover 1.0 rad/sover 0.3 rad/s
  • Camera–IMU report, motion and excitation: the angular velocity of the solved trajectory over time. Gaps and one-axis stretches are easy to see.
  • mvis verdict: grades a result green, yellow or red. It flags a χ² per degree of freedom (the weighted squared error divided by its degrees of freedom, about 1 when the noise model fits) over 1.5 (yellow) or 3.0 (red), a solve that hit its iteration cap, an IMU scale more than 5 % from 1, and a camera time offset over 50 ms. A wrong value from weak motion can still pass these; the 1σ and a second recording are the stronger checks.
  • Calibrate two or more recordings. A value that moves far more than its 1σ between them was not observed.
  • Run the digital twin. MVIS replays your motion on a simulated copy of your rig with known truth, solves it, and shows which values come back (step 7: digital-twin check).
  • Run the round trip. Correct the data with the result and calibrate again; if the first result was right, the second finds nothing left to correct (step 8: round-trip validation).
If you seeAdd to the motion
lever arm weak, or one component weakrotation about the other axes, with brisk changes of turning speed
td,it_{d,i} moves between recordingschanges of speed and direction; avoid long steady sweeps
readout weak or offfaster tilts about the camera’s x and y axes, board across all rows
t_ga weakshort, sharp reversals
Da\mathbf{D}_a weaktilt the rig so each axis points up and down at some point
Dw\mathbf{D}_w weakturns about every axis, at several speeds
distortion differsmove the board into the image corners and edges

The motion in step 2: record the dataset covers all of these in about 60 s. Its motions that fail list the same cases from the recording side.