Skip to content

Notation & glossary

Use this page to look up a symbol, a frame or a sign convention. Every equation on this site uses the symbols below.

The frame and transform conventions follow OpenVINS. Result files use their own XtoY names (below).

A frame is a coordinate system attached to the world, a sensor or the board. We write it in braces, e.g. {I}\{I\}.

SymbolFrameNotes
{G}\{G\}globalgravity-aligned; yaw and position are unobservable
{I}\{I\}base IMUthe reference clock and the trajectory’s body frame
{w}\{w\}, {a}\{a\}gyroscope triad, accelerometer triadinside one IMU; which one {I}\{I\} coincides with depends on the IMU convention
{Isj}\{I_{s_j}\}auxiliary IMU jjrigidly attached to {I}\{I\}
{Ci}\{C_i\}camera iioptical frame: xx right, yy down, zz forward
{B}\{B\}calibration boardfixed in {G}\{G\}

A rotation is written BAR{}^{A}_{B}\mathbf{R}. It takes a vector expressed in {B}\{B\} into {A}\{A\}. A translation ApB{}^{A}\mathbf{p}_{B} is the origin of {B}\{B\} expressed in {A}\{A\}. Together they form the transform BAT{}^{A}_{B}\mathbf{T}:

Ap=BAR Bp+ApB{}^{A}\mathbf{p} = {}^{A}_{B}\mathbf{R}\, {}^{B}\mathbf{p} + {}^{A}\mathbf{p}_{B}

We report a rotation as its rotation vector θ=Log(R)∈R3\boldsymbol{\theta} = \mathrm{Log}(\mathbf{R}) \in \mathbb{R}^3, in rad. Its direction is the rotation axis and its norm is the angle. We never report roll, pitch and yaw angles.

SymbolMeaning
θ∧\boldsymbol{\theta}^\wedgethe 3×33\times3 skew-symmetric matrix of θ\boldsymbol{\theta}, so θ∧v=θ×v\boldsymbol{\theta}^\wedge \mathbf{v} = \boldsymbol{\theta} \times \mathbf{v}
Exp(θ)\mathrm{Exp}(\boldsymbol{\theta})rotation vector to rotation matrix: I+sin⁡θθθ∧+1−cos⁡θθ2(θ∧)2\mathbf{I} + \frac{\sin\theta}{\theta}\boldsymbol{\theta}^\wedge + \frac{1-\cos\theta}{\theta^2}(\boldsymbol{\theta}^\wedge)^2, with θ=∥θ∥\theta = \lVert\boldsymbol{\theta}\rVert
Log(R)\mathrm{Log}(\mathbf{R})the inverse of Exp\mathrm{Exp}; the rotvec fields of a result

The calibration state is the set of quantities we estimate. Intrinsics describe a sensor itself; extrinsics give its pose in the base-IMU frame {I}\{I\}. The Per column says whether a quantity exists once per camera, per rolling-shutter (RS) camera, per IMU or per auxiliary IMU.

SymbolQuantityDimensionPer
fi=[fu fv cu cv]\mathbf{f}_i = [f_u\ f_v\ c_u\ c_v]camera intrinsics (focal lengths, principal point; fx fy cx cy in files)4camera
di\mathbf{d}_idistortion coefficients1 to 5, by camera modelcamera
CiIT{}^{I}_{C_i}\mathbf{T}camera extrinsics6camera
td,it_{d,i}camera–IMU time offset1camera
tr,it_{r,i}rolling-shutter readout1RS camera
Dw\mathbf{D}_wgyroscope correction: scale and misalignment6, or 9 in calyxIMU
Da\mathbf{D}_aaccelerometer correction: scale and misalignment6IMU
wIR{}^{I}_{w}\mathbf{R} or aIR{}^{I}_{a}\mathbf{R}rotation between the gyro and accel triads3 (kalibr, rpng)IMU
Tg\mathbf{T}_gg-sensitivity: how specific force leaks into the gyro9IMU
tgat_{ga}gyro–accel time offset inside one IMU1IMU
IsjIT{}^{I}_{I_{s_j}}\mathbf{T}auxiliary IMU extrinsics6aux IMU
td,jt_{d,j}IMU–IMU time offset1aux IMU
bg, ba\mathbf{b}_g,\ \mathbf{b}_agyro / accel biases3 eachIMU
hkh_kheight of board corner kk along the board normal1board corner, with board_points on

Dw\mathbf{D}_w, Da\mathbf{D}_a and Tg\mathbf{T}_g are correction matrices: they map a raw reading to the true value. IMU model gives the equations.

The full state x\mathbf{x} stacks the trajectory with all of the above. The cost that acts on it is in Calibration graph.

What the sensors report, and the terms we build from those readings:

SymbolMeasurement
ωm\boldsymbol{\omega}_mmeasured angular velocity (raw gyro reading)
am\mathbf{a}_mmeasured specific force (raw accel reading)
Iω{}^{I}\boldsymbol{\omega}true angular velocity of {I}\{I\}, expressed in {I}\{I\}
Ia{}^{I}\mathbf{a}true specific force in {I}\{I\}: Ia=GIR (GaI−Gg){}^{I}\mathbf{a} = {}^{I}_{G}\mathbf{R}\,({}^{G}\mathbf{a}_I - {}^{G}\mathbf{g})
Gg{}^{G}\mathbf{g}gravity in {G}\{G\}, [0 0 −9.81]⊤[0\ 0\ {-9.81}]^\top m/s²
ng, na\mathbf{n}_g,\ \mathbf{n}_agyro / accel white noise (densities sigma_g, sigma_a)
nbg, nba\mathbf{n}_{bg},\ \mathbf{n}_{ba}bias random-walk noise (densities sigma_bg, sigma_ba)
zi,k\mathbf{z}_{i,k}pixel observation of board corner kk in camera ii
Π(⋅)\Pi(\cdot)camera projection: point in {Ci}\{C_i\} to pixel (Camera model)
r\mathbf{r}residual (subscripted by factor type)
Σ\boldsymbol{\Sigma}measurement covariance
Pk\mathbf{P}_kcovariance of the IMU preintegration over interval kk (Base IMU factor)

Every stamp is mapped onto one reference clock, the base-IMU clock.

SymbolMeaning
tIt_Itimestamp on the base-IMU clock, the reference
tCt_Ctimestamp as reported by a camera driver
tdt_doffset such that tI=tC+tdt_I = t_C + t_d
trt_rwhole-frame readout time of a rolling-shutter camera
η(v)\eta(v)row factor of image row vv: the share of trt_r after the frame stamp (Rolling shutter)
tgat_{ga}gyro–accel offset: the accel reading stamped tt was sampled at t−tgat - t_{ga}

A board corner seen in row vv of camera ii was captured at

tI=tC+td,i+η(v) tr,it_I = t_C + t_{d,i} + \eta(v)\, t_{r,i}

The sign convention matters. td is added to the sensor’s own stamp to reach base-IMU time. A positive td means the camera stamps run early: the image was captured later than its stamp says. mvis apply-calib re-stamps each message as stamp + td.

  • Quaternions: result files store rotations as matrices and log(R) rotation vectors, not quaternions. Simulator trajectory files (mvis.simulation) use JPL quaternions qGtoIq_{GtoI}, as in OpenVINS. The Basalt export uses Hamilton quaternions, as Basalt does.
  • Error states (the small corrections around the current estimate) are defined:
    • multiplicatively for rotations, R≈R^ Exp(δθ)\mathbf{R} \approx \hat{\mathbf{R}}\,\mathrm{Exp}(\delta\boldsymbol{\theta});
    • additively for everything else.
  • Units in result files: metres, radians and seconds; summaries add _mm and _ms copies where the name says so.
  • Board points are flat and known by default (estimate.board_points: false). Board errors then go into the calibration: a wrong tag_size gives a scale bias on translations, and a bent board biases the focal length, gyro scale and lever (Why calibrate the board shape). With estimate.board_points: true, MVIS solves one height hkh_k per corner.
  • Factor: one term of the cost the solver minimises.
  • Residual: the error a factor measures, predicted minus observed.
  • Jacobian: the derivative of a residual with respect to a state variable.
  • Lever arm: the translation between two sensors, e.g. IpCi{}^{I}\mathbf{p}_{C_i}.
  • Incidence angle: the angle of a ray off the camera’s optical axis.
  • Specific force: what an accelerometer measures, acceleration minus gravity.
  • Pre-integration: IMU readings between two states combined into one relative motion.
  • Base / auxiliary IMU: the base IMU sets the clock and the body frame; every other IMU is auxiliary.