Results — repeatability across recordings

Batch 2026-11-08a · results commit e893730 · records in calibench-vi/results · the Space shows the same parameter tables.

This page shows repeatability, not accuracy. Repeatability is how much each tool’s estimates scatter when we calibrate the same rig from different recordings.

No ground truth: these datasets have none, so no tool is the reference. A tool can repeat well and still be biased.
Shared recordings only: we compare tools only on recordings that every compared tool ran.
No reprojection error: it measures how well a model fits its own data, so we do not use it here.
Who ran it: the MVIS authors run this board and made every run, including the Kalibr and Basalt runs.

Run dates and data settings

2026-11-08: MVIS and Kalibr on T6 and T11, and Basalt on T11, were re-run with IMU noise measured at rest (white-noise density × 2.5, the same values for every tool; bias walks at the defaults).
2026-11-08: MVIS on T2 and T12 was re-run with the current MVIS code.
2026-11-07: Kalibr camera + IMU on T6 and T11 was re-run from its own camera calibration.
2026-11-07: on T3, every tool ran on the same view-filtered bags (≤ 85°).
2026-11-04: Kalibr on T1 bag 09 was re-run.
2026-11-02: Kalibr on T7 was re-run.
2026-10-22 or earlier: all other Kalibr and Basalt runs.
T6c: the MVIS values come from camera-only runs.

Data settings still differ from MVIS on T6, T7 and T8.

Overall

Real recordings have no ground truth. Measured points are repeatability only: a tool can repeat well and still be biased.

Measured tier A

MVIS 0.78
Kalibr 0.38
Basalt 0.68

3 of 6 aspects have points

Capability tier C

MVIS 0.68
Kalibr 0.51
Basalt 0.43

11 of 11 aspects have points

Usability tier U

MVIS 0.67
Kalibr 0.79
Basalt 0.54

4 of 4 aspects have points

Project facts no points

Fact	MVIS	Kalibr	Basalt
Public since, GitHub stars	unreleased (0.1.0)	2014, about 5.8k	2019, about 1.0k
Peer-reviewed papers	IJRR 2024, on the C++ predecessor	5 papers	3DV 2018, CVPR 2020
Licence	GPL-3.0 or commercial	BSD	BSD-3

Licence: BSD: use in closed products with the notice kept. GPL-3.0: free, but products that ship it must be GPL too; a commercial licence is available.

Derived total derived

MVIS 0.71
Kalibr 0.56
Basalt 0.55

The derived total is the mean of the three tier values, each tier the mean of its aspects. Sum of points over the 18 aspects that have points: MVIS 12.45, Kalibr 9.89, Basalt 8.99.

Maturity, papers and licence are shown, not counted. Counting them as in the first draft puts Basalt first (0.67 against 0.65 for MVIS).

Equal tiers: MVIS first in 18 of 18 leave-one-aspect-out variants.

Head to head (measured)

Camera-only repeatability: MVIS and Basalt repeat about equally well; both repeat better than Kalibr.

Camera-only repeatability: points per pair of tools and family
Pair	Families (points)	Points
MVIS vs Kalibr	T1 1.00 / 0.00, T3 0.67 / 0.33, T4 0.33 / 0.67, T6c 1.00 / 0.00	0.75 / 0.25
MVIS vs Basalt	T3 0.50 / 0.50, T4 0.50 / 0.50	0.50 / 0.50
Kalibr vs Basalt	T3 0.17 / 0.83, T4 0.67 / 0.33	0.42 / 0.58

Camera-IMU repeatability: MVIS repeats better than Kalibr (T6, T11); about equal: MVIS and Basalt (T11), Kalibr and Basalt (T11).

Camera-IMU repeatability: points per pair of tools and family
Pair	Families (points)	Points
MVIS vs Kalibr	T6 1.00 / 0.00, T11 0.50 / 0.50	0.75 / 0.25
MVIS vs Basalt	T11 0.50 / 0.50	0.50 / 0.50
Kalibr vs Basalt	T11 0.50 / 0.50	0.50 / 0.50

Aspects

● 1 · ◐ ½ · ○ 0; other values are means of sub-checks, which open under the aspect name. – = no points, with the reason.

Measured · 3 of 6 aspects have points
Aspect	Home	MVIS	Kalibr	Basalt
A1 Camera-only repeatability Rule Spread of focal length, principal point, distortion and camera distance across recordings of one rig.	–	0.676 pair-families	0.316 pair-families	0.544 pair-families
A2 Camera-IMU repeatability Rule Spread of the camera-IMU distance, gyro and accel terms. The time offset is left out: MVIS also fits a gyro-accel delay.	–	0.673 pair-families	0.333 pair-families	0.502 pair-families
A3 Rolling-shutter readout repeatability Rule Readout spread. On T2 the lens models differ (MVIS equidistant, Kalibr RS radtan); T8 and T11 have one tool only.	–	–inputs differ between tools	–inputs differ between tools	–inputs differ between tools
A4 Accuracy against known truth Rule Error against simulation truth, all tools on the same rendered data. Accuracy against known truth: not measured for any tool. It needs simulated recordings run through all three tools.	–	–not measured	–not measured	–not measured
A5 Solve robustness Rule Share of attempted rows that end ok and physical, on rows the tool's model supports. 1 = at least 99 %, 0.5 = 90-99 %, 0 = under 90 %. Kalibr: one final fail and one unphysical IMU fit. Basalt: our build with robustness patches; physicality on three families unchecked.	–	1.00176 of 176	0.5090 of 92	1.0076 of 76, provisional
A6 Runtime Rule Bag to calibration, detection included, same machine, same rows. Timing definitions differ between the tools today.	–	–pending	–pending	–pending

Capability · 11 of 11 aspects have points

Aspect

Home

MVIS

Kalibr

Basalt

C1 Lens models, fisheye included

Rule

(a) pinhole-radtan (b) KB4 / equidistant (c) a unified wide model (d) mixed models per rig

Shared

1.00

C2 Calibration targets

Rule

1 = two or more target families, 0.5 = one

Shared

1.00

0.50AprilGrid only

C3 Board model

Rule

Estimates a non-planar board

Designed into MVIS

1.00

0.00

C4 Rolling shutter

Rule and sub-checks

Sub-check	MVIS	Kalibr
RS single camera	1	1
RS multi-camera	1	0
RS camera + IMU	1	0
Mixed GS + RS rig with IMU	1	0

Designed into MVIS

1.00

0.25

0.00

C5 Time offsets and clocks

Rule and sub-checks

Sub-check	MVIS	Kalibr	Basalt
Camera-IMU time offset	1	1	1
Per-camera time offset from a free start (0.5 = needs roughly synced cameras)	1	0.5	0
Camera-camera offset without an IMU on the default path (0.5 = exists, not on the command-line path)	0.5	0	0
Host-clock sync for network-stamped sensors (Kalibr feature)	0	1	0
Gyro-accel time offset (Designed into MVIS)	1	0	0

Mixed

0.70

0.50

0.20

C6 IMU intrinsic model

Rule and sub-checks

Sub-check	MVIS	Kalibr	Basalt
Scale + misalignment, gyro and accel	1	1	1
Gyro g-sensitivity	1	1	0
Accel size effect (Kalibr feature)	0	1	0
Bias random walk	1	1	0

Mixed

0.75

1.00

0.25

C7 Multi-sensor rigs

Rule and sub-checks

Sub-check	MVIS	Kalibr	Basalt
Several cameras + IMU	1	1	1
Several IMUs	1	1	0
Several IMUs (Designed into MVIS)	1	0.5	0

Mixed

1.00

0.83

0.33

C8 Trajectory model

Rule

1 = continuous-time spline; 0.5 = discrete states with a known bias

Kalibr feature / Basalt feature

0.50

1.00

C9 Mocap-to-IMU

Rule

Supported

Basalt feature

0.00

1.00

C10 Vignetting

Rule

Experimental = 0.5

Basalt feature

0.00

0.50

C11 Calibration simulation

Rule

1 = scenario generator with images or bags; 0.5 = re-simulates corners from a calibration

Designed into MVIS

0.50

0.00

Usability · 4 of 4 aspects have points

Aspect

Home

MVIS

Kalibr

Basalt

U1 Install and platforms

Rule

1 = prebuilt packages for a current OS; 0.5 = source build or container

Shared

0.50

1.00

U2 Public docs per feature

Rule

1 = a guide or wiki page per feature; 0.5 = one calibration doc

Shared

1.00

0.50

U3 Reports and diagnostics

Rule and sub-checks

Sub-check	MVIS	Kalibr	Basalt
Per-run report (0.5 = GUI view only)	1	1	0.5
Apply a calibration and recalibrate in the tool	1	0	0
Standalone validator or focus tool	0	1	0

Shared

0.67

0.17

U4 Export formats

Rule

1 = two or more external formats or tools; 0.5 = one

Shared

0.50

1.00

0.50

Reader presets

The derived total under other weightings; no preset replaces the tier cards
Preset	MVIS	Kalibr	Basalt	Leads
Equal tiers A, C, U (default)	0.71	0.56	0.55	MVIS
Algorithm only (A + C)	0.73	0.44	0.56	MVIS
Capability only	0.68	0.51	0.43	MVIS
Usability only	0.67	0.79	0.54	Kalibr
Capability without any tool's home rows or sub-checks	0.97	0.90	0.53	MVIS
Measured only (A1 + A2 + A5)	0.78	0.38	0.68	MVIS
Camera-only measured (A1)	0.67	0.31	0.54	MVIS
Camera-only user (C1-C3, C4 a-b, C10, C11, A1)	0.74	0.40	0.36	MVIS
GS camera-IMU user (C5-C9)	0.59	0.67	0.56	Kalibr
RS or mixed-rig camera-IMU user (C4-C9)	0.66	0.60	0.46	MVIS
First draft: maturity, papers, licence counted (old tier G)	0.65	0.61	0.67	Basalt

Columns, rules and what has no points

Kalibr, default settings (view shuffle on). Camera-only runs start from the MVIS focal length.
Basalt 0.1.7. Capabilities as released; measured cells from our headless build with robustness patches.
T6 counted despite 'settings differ': Kalibr now starts from its own camera calibration; the remaining differences (board shape, data windows) are tool settings.
T7 left out: Kalibr has no rolling-shutter camera-IMU model, so no like-for-like T7 run exists.
IMU-only calibration without a camera: none of the three tools, so it does not separate them.
Number of IMU conventions: a convention, not a quality lever.
Uncertainty output: pending, Basalt not checked.
Determinism: MVIS gives the same result when run twice (to rounding); Kalibr's default view shuffle varies run to run. Not counted until it has a rule.

VI-Rig: 4 IMUs + 3 cameras

Our flagship rig. Dataset page.

VI-Rig 3 cams: which tool repeats more tightly, per family and group of parameters
Family	Focal + pp px	Distortion	Distance mm	Time offset ms	Cam–IMU distance mm	Gyro	Accel
T6c · camera n = 10 matched 3 GS cameras — camera calibration	MVIS = Basalt> Kalibr ×4.6 ^ab	MVIS = Basalt> Kalibr ×9.9 ^cab	Kalibr only	—	—	—	—
T6 · cam+IMU n = 10 matched 3 GS cameras + 4 IMUs	input, not compared	input, not compared	input, not compared	Tie ^de	MVIS ×3.4 ^d	MVIS ×2.5K-IMU ^df	MVIS ×2.7K-IMU ^df

T6c · camera · 3 GS cameras — camera calibration More repeatable or tie: Focal + pp MVIS = Basalt > Kalibr · Distortion MVIS = Basalt > Kalibr n = 10 matched · MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10

Focal + pp MVIS= Basalt > Kalibr×4.6^ab
Distortion MVIS= Basalt > Kalibr×9.9^cab
Distance: Kalibr only

T6 · cam+IMU · 3 GS cameras + 4 IMUs More repeatable or tie: Time offset tie · Cam–IMU distance MVIS · Gyro MVIS K-IMU · Accel MVIS K-IMU n = 10 matched · MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run

Time offsetTie^de
Cam–IMU distance MVIS×3.4^d
Gyro MVIS×2.5K-IMU^df
Accel MVIS×2.7K-IMU^df
Focal + pp · Distortion · Distance: input, not compared

MVIS Kalibr Basalt tool ×k more repeatable: the other tools' typical scatter is k times larger Tie gap within chance or under 20 % too few under 6 shared recordings input no comparison, reason in the cell K-IMU MVIS run with Kalibr's IMU model

k = the looser tool's typical scatter divided by the tighter tool's (median over the group's parameters, on the recordings every compared tool ran). Click a cell for its numbers.

a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).
b The tool pairs here use different parameters, because one tool is not compared on every camera; the cell lists which.
c Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
d Data settings still differ between the tools on this family.
e MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.
f MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Each cell is a separate comparison on the recordings every compared tool ran. Cells are not added up and nothing is ordered across datasets.

Not part of this comparison

T6 reprojection error (3 rows): fit quality, not repeatability and not accuracy
T6 IMU g-sensitivity Tg (36 rows): Kalibr absorbs it in its IMU model
T6 gyro–accel delay t_ga (4 rows): MVIS only
T6 Basalt: not run: one IMU per run; no run holds the whole rig

Focal length + principal point · T6c · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3 GS cameras — camera calibration, camera

The spreads of MVIS and Basalt differ by 1.00×, but a gap this size appears often by chance with 10 recordings (p = 0.99). Treated as a tie. Both scatter less than Kalibr: across the same 10 recordings Kalibr's estimates scatter about 4.6× more than MVIS's and 5.1× more than Basalt's. Holds with any one recording left out.

MVIS here: the camera values of the T6 camera+IMU solve.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 12 param.	0.22	0.13 – 0.39	0.0059 (raw 0.002)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	1	0.83 – 1.74	0.99	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	5.07	2.89 – 14.33	0.0059 (raw 0.002)	Basalt tighter	holds with any one recording left out

Ratio R of the two tools' typical scatter, first ÷ second. Shaded: 0.80–1.25, the gap treated as a tie even when it is clear. Bars: 95 % interval (not adjusted for multiple tests); a missing end is not bounded. Filled dots in a tool's colour mark a decided pair.

Parameter	MVIS sd	Kalibr sd	Basalt sd
Blackfly · fx px	0.983	2.17	–
Blackfly · fy px	1.12	2.5	–
Blackfly · cx px	0.551	3.28	–
Blackfly · cy px	0.661	1.44	–
t265 fisheye 1 · fx px	0.392	1.61	0.403
t265 fisheye 1 · fy px	0.316	1.69	0.332
t265 fisheye 1 · cx px	0.205	0.553	0.204
t265 fisheye 1 · cy px	0.207	1.09	0.175
t265 fisheye 2 · fx px	0.395	1.99	0.395
t265 fisheye 2 · fy px	0.317	2.08	0.333
t265 fisheye 2 · cx px	0.217	0.589	0.204
t265 fisheye 2 · cy px	0.164	1.01	0.135

Not in this comparison

Basalt · shown only · 4 rows: other camera model (pinhole-radtan8)

a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).
b The tool pairs here use different parameters, because one tool is not compared on every camera; the cell lists which.

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T6c · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3 GS cameras — camera calibration, camera

The spreads of MVIS and Basalt differ by 1.5×, but a gap this size appears often by chance with 10 recordings (p = 0.055). Treated as a tie. Both scatter less than Kalibr: across the same 10 recordings Kalibr's estimates scatter about 9.9× more than MVIS's and 19× more than Basalt's. Holds with any one recording left out.

MVIS here: the camera values of the T6 camera+IMU solve.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 12 param.	0.1	0.05 – 0.32	0.0059 (raw 0.002)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	1.47	1 – 2.74	0.055	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	19.49	4.54 – 66.05	0.0059 (raw 0.002)	Basalt tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Basalt sd
Blackfly · dist k1	0.00256	0.00996	–
Blackfly · dist k2	0.00261	0.00889	–
Blackfly · dist p1	0.000167	0.000385	–
Blackfly · dist p2	0.000185	0.000954	–
t265 fisheye 1 · dist k1	0.00155	0.0188	0.00115
t265 fisheye 1 · dist k2	0.0037	0.042	0.00248
t265 fisheye 1 · dist k3	0.00435	0.0373	0.00239
t265 fisheye 1 · dist k4	0.0017	0.011	0.000761
t265 fisheye 2 · dist k1	0.0013	0.0324	0.00104
t265 fisheye 2 · dist k2	0.00336	0.0785	0.00276
t265 fisheye 2 · dist k3	0.00419	0.073	0.00289
t265 fisheye 2 · dist k4	0.00171	0.0224	0.001

Not in this comparison

Basalt · shown only · 4 rows: other camera model (pinhole-radtan8)

c Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).
b The tool pairs here use different parameters, because one tool is not compared on every camera; the cell lists which.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T6c · Kalibr only (MVIS no value · Kalibr 10 of 10 · Basalt no value) · 3 GS cameras — camera calibration, camera

Kalibr only: no value for this tool

Not in this comparison

MVIS · missing · 2 rows: no value for this tool
Basalt · missing · 2 rows: not in one run

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T6 · Input, not compared (MVIS 10 of 10 · Kalibr input · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 12 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 12 rows: not run: one IMU per run; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T6 · Input, not compared (MVIS 10 of 10 · Kalibr input · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 12 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 12 rows: not run: one IMU per run; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T6 · Input, not compared (MVIS 10 of 10 · Kalibr input · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 2 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 2 rows: not run: one IMU per run; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T6 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

The spreads of MVIS and Kalibr differ by 1.03×, but a gap this size appears often by chance with 10 recordings (p = 0.73). Treated as a tie.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 6 param.	1.03	0.82 – 1.33	0.73	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Ratio
Blackfly · time offset td to base IMU ms	0.0964	0.149	0.65
t265 fisheye 1 · time offset td to base IMU ms	0.155	0.247	0.63
t265 fisheye 2 · time offset td to base IMU ms	0.237	0.244	0.97
GX3-25 · time delay td to base IMU ms	0.208	0.189	1.10
GX3-35 · time delay td to base IMU ms	0.209	0.17	1.23
t265 IMU · time delay td to base IMU ms	0.212	0.188	1.13

Not in this comparison

Basalt · 6 rows: not run: one IMU per run; no run holds the whole rig

d Data settings still differ between the tools on this family.
e MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T6 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Across the same 10 recordings, Kalibr's estimates scatter about 3.4× more than MVIS's (median over 6 parameters). A gap this large appeared in 0.59 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 6 param.	0.29	0.16 – 0.51	0.0059	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
Blackfly · distance to base IMU ‖p‖ mm	1.41	3.64	0.39
t265 fisheye 1 · distance to base IMU ‖p‖ mm	1.48	3.19	0.46
t265 fisheye 2 · distance to base IMU ‖p‖ mm	1.7	3.99	0.43
GX3-25 · distance to base IMU ‖p‖ mm	0.198	0.883	0.22
GX3-35 · distance to base IMU ‖p‖ mm	0.189	0.862	0.22
t265 IMU · distance to base IMU ‖p‖ mm	0.194	1.42	0.14

Not in this comparison

Basalt · 6 rows: not run: one IMU per run; no run holds the whole rig

d Data settings still differ between the tools on this family.

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T6 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Across the same 10 recordings, Kalibr's estimates scatter about 2.5× more than MVIS's (median over 24 parameters). A gap this large appeared in 0.59 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 24 param.	0.4	0.25 – 0.6	0.0059	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
Xsens MTi · gyro Dw[0,0] scale x	0.00149	0.00344	0.43
Xsens MTi · gyro Dw[1,1] scale y	0.00051	0.00182	0.28
Xsens MTi · gyro Dw[2,2] scale z	0.00201	0.00248	0.81
Xsens MTi · gyro Dw[1,0] cross	0.000808	0.00266	0.30
Xsens MTi · gyro Dw[2,0] cross	0.00143	0.00393	0.36
Xsens MTi · gyro Dw[2,1] cross	0.00144	0.00345	0.42
GX3-25 · gyro Dw[0,0] scale x	0.000544	0.00193	0.28
GX3-25 · gyro Dw[1,1] scale y	0.00161	0.0039	0.41
GX3-25 · gyro Dw[2,2] scale z	0.00235	0.00197	1.19
GX3-25 · gyro Dw[1,0] cross	0.000649	0.00239	0.27
GX3-25 · gyro Dw[2,0] cross	0.00149	0.00333	0.45
GX3-25 · gyro Dw[2,1] cross	0.0016	0.00438	0.37
GX3-35 · gyro Dw[0,0] scale x	0.000392	0.00172	0.23
GX3-35 · gyro Dw[1,1] scale y	0.0016	0.00388	0.41
GX3-35 · gyro Dw[2,2] scale z	0.00228	0.00211	1.08
GX3-35 · gyro Dw[1,0] cross	0.000785	0.00242	0.32
GX3-35 · gyro Dw[2,0] cross	0.00131	0.0032	0.41
GX3-35 · gyro Dw[2,1] cross	0.00175	0.00448	0.39
t265 IMU · gyro Dw[0,0] scale x	0.00139	0.00344	0.40
t265 IMU · gyro Dw[1,1] scale y	0.00203	0.00276	0.74
t265 IMU · gyro Dw[2,2] scale z	0.000568	0.00208	0.27
t265 IMU · gyro Dw[1,0] cross	0.00163	0.00408	0.40
t265 IMU · gyro Dw[2,0] cross	0.000649	0.00264	0.25
t265 IMU · gyro Dw[2,1] cross	0.00141	0.00351	0.40

Not in this comparison

Basalt · 24 rows: not run: one IMU per run; no run holds the whole rig

d Data settings still differ between the tools on this family.
f MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T6 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run) · 3 GS cameras + 4 IMUs, cam+IMU

Across the same 10 recordings, Kalibr's estimates scatter about 2.7× more than MVIS's (median over 12 parameters). A gap this large appeared in 0.2 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 12 param.	0.37	0.27 – 0.58	0.002	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
Xsens MTi · accel Da[0,0] scale x	0.000437	0.000925	0.47
Xsens MTi · accel Da[1,1] scale y	0.00163	0.00332	0.49
Xsens MTi · accel Da[2,2] scale z	0.00132	0.00292	0.45
GX3-25 · accel Da[0,0] scale x	0.00179	0.00627	0.29
GX3-25 · accel Da[1,1] scale y	0.000409	0.00167	0.24
GX3-25 · accel Da[2,2] scale z	0.00137	0.00239	0.57
GX3-35 · accel Da[0,0] scale x	0.00173	0.0053	0.33
GX3-35 · accel Da[1,1] scale y	0.000384	0.00149	0.26
GX3-35 · accel Da[2,2] scale z	0.00133	0.00315	0.42
t265 IMU · accel Da[0,0] scale x	0.000525	0.00611	0.09
t265 IMU · accel Da[1,1] scale y	0.0016	0.00544	0.29
t265 IMU · accel Da[2,2] scale z	0.002	0.00206	0.97

Not in this comparison

Basalt · 12 rows: not run: one IMU per run; no run holds the whole rig

d Data settings still differ between the tools on this family.
f MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Parameter tables, tool differences and run configs (VI-Rig 3 cams)

No ground truth: ◆ marks disagreement, not who is right.

median ± sd (recordings) · (3/15) partial · fixed Kalibr input · shown / other model, no Δ · ≠ model, fit: no Δ · ◆ |Δ| above the larger sd · ≠n different recording counts · ‡ see How to read this

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

Key and MVIS column

Swipe a table sideways for the other tools and differences.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

T6c · 3 GS cameras — camera calibration

MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10, Blackfly shown only

Camera · april 01–10 · Kalibr kalibr_calibrate_cameras, 3 cameras jointly

Read first: MVIS runs the Blackfly and the t265 pair separately, so it has no baseline to the Blackfly; Kalibr solves the three cameras jointly.

Notes

Like for like: MVIS uses two corners per tag and estimates the board shape; Kalibr uses all four corners and assumes a flat board.

Note: MVIS runs camera-only per bag, the Blackfly alone and the t265 pair (the rig's cameras share no frame stamps), so it has no baseline to the Blackfly; Kalibr's 10 camera-only runs solve the three cameras jointly.

Basalt: t265 pair (kb4, fisheye 2 snapped onto fisheye 1) and Blackfly alone (pinhole-radtan8, a different model from radtan: its intrinsics are shown only); no camera-only Basalt run holds both devices, so no baselines to the Blackfly.

cam0 – Blackfly · 8 rows · ◆ 3

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr
fx px	449.95 ± 0.98 (10)	446.75 ± 2.17 (10)	shown 448.12 ± 1.02 (10)	◆ +3.20
fy px	449.10 ± 1.12 (10)	444.89 ± 2.50 (10)	shown 446.93 ± 0.98 (10)	◆ +4.21
cx px	369.81 ± 0.55 (10)	368.67 ± 3.28 (10)	shown 368.56 ± 0.47 (10)	+1.13
cy px	279.90 ± 0.66 (10)	278.41 ± 1.44 (10)	shown 281.06 ± 0.51 (10)	◆ +1.50
dist k1	−0.31405 ± 0.00256 (10)	−0.30796 ± 0.00996 (10)	shown 0.31003 ± 6.46473 (10)	−0.00609
dist k2	0.084 ± 0.003 (10)	0.077 ± 0.009 (10)	shown 11.204 ± 13.707 (10)	+0.007
dist p1	0.00003 ± 0.00017 (10)	0.00011 ± 0.00038 (10)	shown −0.00034 ± 0.00026 (10)	−0.00008
dist p2	−0.00013 ± 0.00018 (10)	0.00014 ± 0.00095 (10)	shown −0.00032 ± 0.00024 (10)	−0.00027

cam1 – t265 fisheye 1 · 9 rows · ◆ 5

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	286.30 ± 0.39 (10)	285.72 ± 1.61 (10)	286.17 ± 0.40 (10)	+0.58	+0.45	−0.13
fy px	285.95 ± 0.32 (10)	285.13 ± 1.69 (10)	285.60 ± 0.33 (10)	+0.82	+0.47	◆ −0.35
cx px	415.16 ± 0.20 (10)	415.83 ± 0.55 (10)	415.00 ± 0.20 (10)	◆ −0.67	◆ −0.83	−0.16
cy px	403.99 ± 0.21 (10)	404.09 ± 1.09 (10)	404.05 ± 0.17 (10)	−0.10	−0.05	+0.06
dist k1	−0.00451 ± 0.00155 (10)	0.00165 ± 0.01882 (10)	−0.00455 ± 0.00115 (10)	−0.00617	−0.00621	−0.00004
dist k2	0.03678 ± 0.00370 (10)	0.02307 ± 0.04201 (10)	0.04107 ± 0.00248 (10)	+0.01371	+0.01800	◆ +0.00429
dist k3	−0.03395 ± 0.00435 (10)	−0.02549 ± 0.03728 (10)	−0.04057 ± 0.00239 (10)	−0.00847	−0.01508	◆ −0.00662
dist k4	0.00501 ± 0.00170 (10)	0.00321 ± 0.01100 (10)	0.00755 ± 0.00076 (10)	+0.00180	+0.00434	◆ +0.00254
baseline to Blackfly mm	not in one run	71.856 ± 1.457 (10)	not in one run

cam2 – t265 fisheye 2 · 9 rows · ◆ 3

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	286.28 ± 0.39 (10)	285.37 ± 1.99 (10)	286.05 ± 0.39 (10)	+0.91	+0.68	−0.23
fy px	285.73 ± 0.32 (10)	284.57 ± 2.08 (10)	285.32 ± 0.33 (10)	+1.15	+0.75	◆ −0.40
cx px	414.58 ± 0.22 (10)	415.44 ± 0.59 (10)	414.58 ± 0.20 (10)	◆ −0.86	◆ −0.86	0.00
cy px	398.45 ± 0.16 (10)	398.62 ± 1.01 (10)	398.38 ± 0.13 (10)	−0.16	−0.24	−0.08
dist k1	−0.00618 ± 0.00130 (10)	0.00218 ± 0.03237 (10)	−0.00554 ± 0.00104 (10)	−0.00837	−0.00773	+0.00064
dist k2	0.04011 ± 0.00336 (10)	0.02374 ± 0.07849 (10)	0.04308 ± 0.00276 (10)	+0.01637	+0.01934	+0.00297
dist k3	−0.03668 ± 0.00419 (10)	−0.02631 ± 0.07301 (10)	−0.04154 ± 0.00289 (10)	−0.01037	−0.01523	◆ −0.00486
dist k4	0.00632 ± 0.00171 (10)	0.00280 ± 0.02243 (10)	0.00763 ± 0.00100 (10)	+0.00352	+0.00483	+0.00131
baseline to Blackfly mm	not in one run	71.630 ± 2.455 (10)	not in one run

T6 · 3 GS cameras + 4 IMUs

MVIS 10 of 10 · Kalibr 10 of 10 · Basalt not run

Camera + IMU · april 01–10 · Kalibr 3 cameras + 4 IMUs jointly

Read first: Kalibr's intrinsics and baselines are inputs from its own camera-only calibration; its td and lever are solved given them. Data settings differ: see Notes.

Notes, including why Basalt has no run

Basalt not run: Basalt cannot calibrate T6 — one IMU per run (the rig has four) and frames grouped by exact stamp (the t265 pair and the Blackfly share none), so no single run holds the rig.

Like for like: MVIS uses two corners per tag, non-base cameras at 4 Hz or lower and an estimated board shape; Kalibr has none of these. Every tool uses the same IMU noise, measured at rest (the rig's noisiest IMU, one set for all four IMUs). Kalibr's runs date from 2026-11-08.

Note: Kalibr starts from its own camera-only calibration of each bag. Every tool uses the same IMU noise, measured at rest on the rig's noisiest IMU.

cam0 – Blackfly · 11 rows · ◆ 1

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
fx px	448.30 ± 0.61 (10)	fixed 446.75 ± 2.17 (10)
fy px	447.62 ± 0.58 (10)	fixed 444.89 ± 2.50 (10)
cx px	370.78 ± 0.71 (10)	fixed 368.67 ± 3.28 (10)
cy px	281.31 ± 0.69 (10)	fixed 278.41 ± 1.44 (10)
dist k1	−0.31178 ± 0.00310 (10)	fixed −0.30796 ± 0.00996 (10)
dist k2	0.08088 ± 0.00268 (10)	fixed 0.07731 ± 0.00889 (10)
dist p1	−0.00033 ± 0.00010 (10)	fixed 0.00011 ± 0.00038 (10)
dist p2	−0.00012 ± 0.00013 (10)	fixed 0.00014 ± 0.00095 (10)
time offset td to base IMU^‡ ms	−4.293 ± 0.093 (10)	−5.438 ± 0.149 (10)	◆ +1.145
distance to base IMU ‖p‖ mm	89.32 ± 1.41 (10)	91.41 ± 3.64 (10)	−2.09
reproj median px	0.349 ± 0.032 (10)	0.483 ± 0.209 (10)	fit

cam1 – t265 fisheye 1 · 12 rows · ◆ 2

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
fx px	287.08 ± 0.36 (10)	fixed 285.72 ± 1.61 (10)
fy px	286.70 ± 0.34 (10)	fixed 285.13 ± 1.69 (10)
cx px	415.32 ± 0.27 (10)	fixed 415.83 ± 0.55 (10)
cy px	403.53 ± 0.25 (10)	fixed 404.09 ± 1.09 (10)
dist k1	−0.00341 ± 0.00286 (10)	fixed 0.00165 ± 0.01882 (10)
dist k2	0.03197 ± 0.00592 (10)	fixed 0.02307 ± 0.04201 (10)
dist k3	−0.03046 ± 0.00596 (10)	fixed −0.02549 ± 0.03728 (10)
dist k4	0.00456 ± 0.00208 (10)	fixed 0.00321 ± 0.01100 (10)
baseline to Blackfly mm	71.267 ± 0.140 (10)	fixed 71.856 ± 1.457 (10)
time offset td to base IMU^‡ ms	−5.022 ± 0.149 (10)	−5.999 ± 0.247 (10)	◆ +0.977
distance to base IMU ‖p‖ mm	130.10 ± 1.48 (10)	133.92 ± 3.19 (10)	◆ −3.83
reproj median px	0.225 ± 0.013 (10)	0.272 ± 0.126 (10)	fit

cam2 – t265 fisheye 2 · 12 rows · ◆ 1

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
fx px	287.02 ± 0.33 (10)	fixed 285.37 ± 1.99 (10)
fy px	286.40 ± 0.32 (10)	fixed 284.57 ± 2.08 (10)
cx px	414.56 ± 0.32 (10)	fixed 415.44 ± 0.59 (10)
cy px	397.96 ± 0.28 (10)	fixed 398.62 ± 1.01 (10)
dist k1	−0.00397 ± 0.00264 (10)	fixed 0.00218 ± 0.03237 (10)
dist k2	0.03443 ± 0.00611 (10)	fixed 0.02374 ± 0.07849 (10)
dist k3	−0.03139 ± 0.00675 (10)	fixed −0.02631 ± 0.07301 (10)
dist k4	0.00430 ± 0.00258 (10)	fixed 0.00280 ± 0.02243 (10)
baseline to Blackfly mm	69.849 ± 0.327 (10)	fixed 71.630 ± 2.455 (10)
time offset td to base IMU^‡ ms	−7.798 ± 0.237 (10)	−8.778 ± 0.244 (10)	◆ +0.980
distance to base IMU ‖p‖ mm	111.39 ± 1.72 (10)	114.03 ± 3.99 (10)	−2.63
reproj median px	0.239 ± 0.014 (10)	0.288 ± 0.140 (10)	fit

base IMU – Xsens MTi · 19 rows · ◆ 4

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	0.99863 ± 0.00149 (10)	0.99998 ± 0.00344 (10)	−0.00135
gyro Dw[1,1] scale y	0.99778 ± 0.00051 (10)	0.99679 ± 0.00182 (10)	+0.00099
gyro Dw[2,2] scale z	1.00183 ± 0.00201 (10)	1.00589 ± 0.00248 (10)	◆ −0.00406
gyro Dw[1,0] cross	−0.00147 ± 0.00081 (10)	0.00235 ± 0.00266 (10)	◆ −0.00382
gyro Dw[2,0] cross	0.00329 ± 0.00143 (10)	0.00682 ± 0.00393 (10)	−0.00353
gyro Dw[2,1] cross	0.00179 ± 0.00144 (10)	0.01213 ± 0.00345 (10)	◆ −0.01034
accel Da[0,0] scale x	0.99988 ± 0.00044 (10)	1.00005 ± 0.00093 (10)	−0.00017
accel Da[1,1] scale y	0.99681 ± 0.00163 (10)	0.99549 ± 0.00332 (10)	+0.00132
accel Da[2,2] scale z	1.00043 ± 0.00132 (10)	1.00476 ± 0.00292 (10)	◆ −0.00433
Tg[0,0] g-sensitivity mrad·s/m	0.016 ± 0.048 (10)	0.512 ± 0.824 (10)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.074 ± 0.087 (10)	0.042 ± 0.564 (10)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	0.053 ± 0.085 (10)	2.760 ± 1.449 (10)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.033 ± 0.033 (10)	0.184 ± 0.967 (10)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.004 ± 0.064 (10)	0.032 ± 0.643 (10)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.020 ± 0.086 (10)	3.582 ± 1.728 (10)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.012 ± 0.074 (10)	0.101 ± 0.275 (10)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	0.058 ± 0.090 (10)	−1.158 ± 0.364 (10)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.100 ± 0.087 (10)	0.093 ± 0.296 (10)	≠ model
t_ga gyro−accel ms	−0.307 ± 0.109 (10)	not in model

aux IMU – GX3-25 · 21 rows · ◆ 4

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	0.99989 ± 0.00054 (10)	0.99849 ± 0.00193 (10)	+0.00140
gyro Dw[1,1] scale y	1.00107 ± 0.00161 (10)	1.00185 ± 0.00390 (10)	−0.00078
gyro Dw[2,2] scale z	1.00310 ± 0.00235 (10)	1.01022 ± 0.00197 (10)	◆ −0.00712
gyro Dw[1,0] cross	−0.00110 ± 0.00065 (10)	0.00045 ± 0.00239 (10)	−0.00155
gyro Dw[2,0] cross	−0.00240 ± 0.00149 (10)	−0.01170 ± 0.00333 (10)	◆ +0.00929
gyro Dw[2,1] cross	0.00170 ± 0.00160 (10)	−0.00122 ± 0.00438 (10)	+0.00293
accel Da[0,0] scale x	0.99729 ± 0.00179 (10)	0.99594 ± 0.00627 (10)	+0.00134
accel Da[1,1] scale y	1.00073 ± 0.00041 (10)	1.00340 ± 0.00167 (10)	◆ −0.00267
accel Da[2,2] scale z	1.00040 ± 0.00137 (10)	1.00156 ± 0.00239 (10)	−0.00116
Tg[0,0] g-sensitivity mrad·s/m	0.041 ± 0.110 (10)	−1.083 ± 0.406 (10)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.038 ± 0.060 (10)	−0.181 ± 0.517 (10)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.072 ± 0.126 (10)	1.498 ± 0.973 (10)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	0.001 ± 0.092 (10)	−0.526 ± 0.339 (10)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.026 ± 0.056 (10)	−0.033 ± 0.398 (10)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.038 ± 0.076 (10)	1.217 ± 0.829 (10)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.036 ± 0.087 (10)	−0.224 ± 0.341 (10)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	0.002 ± 0.067 (10)	0.131 ± 0.245 (10)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.103 ± 0.091 (10)	1.146 ± 0.569 (10)	≠ model
time delay td to base IMU^‡ ms	−0.032 ± 0.209 (10)	−0.133 ± 0.189 (10)	+0.101
distance to base IMU ‖p‖ mm	131.31 ± 0.20 (10)	129.26 ± 0.88 (10)	◆ +2.05
t_ga gyro−accel ms	0.147 ± 0.068 (10)	not in model

aux IMU – GX3-35 · 21 rows · ◆ 6

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	0.99888 ± 0.00039 (10)	0.99818 ± 0.00172 (10)	+0.00070
gyro Dw[1,1] scale y	0.99978 ± 0.00160 (10)	1.00242 ± 0.00388 (10)	−0.00264
gyro Dw[2,2] scale z	1.00149 ± 0.00228 (10)	1.00859 ± 0.00211 (10)	◆ −0.00710
gyro Dw[1,0] cross	−0.00161 ± 0.00078 (10)	0.00090 ± 0.00242 (10)	◆ −0.00250
gyro Dw[2,0] cross	−0.00179 ± 0.00131 (10)	−0.01102 ± 0.00320 (10)	◆ +0.00923
gyro Dw[2,1] cross	−0.00232 ± 0.00175 (10)	−0.00612 ± 0.00448 (10)	+0.00381
accel Da[0,0] scale x	0.99673 ± 0.00173 (10)	0.99543 ± 0.00530 (10)	+0.00130
accel Da[1,1] scale y	1.00005 ± 0.00038 (10)	1.00237 ± 0.00149 (10)	◆ −0.00232
accel Da[2,2] scale z	1.00009 ± 0.00133 (10)	1.00623 ± 0.00315 (10)	◆ −0.00614
Tg[0,0] g-sensitivity mrad·s/m	0.006 ± 0.100 (10)	−0.924 ± 0.474 (10)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	−0.026 ± 0.034 (10)	−0.013 ± 0.691 (10)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.061 ± 0.081 (10)	−0.251 ± 1.086 (10)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.024 ± 0.069 (10)	−0.556 ± 0.390 (10)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.024 ± 0.031 (10)	0.182 ± 0.584 (10)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.038 ± 0.052 (10)	−0.144 ± 1.024 (10)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.042 ± 0.086 (10)	0.140 ± 0.334 (10)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	0.009 ± 0.031 (10)	0.188 ± 0.251 (10)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.012 ± 0.061 (10)	1.292 ± 0.607 (10)	≠ model
time delay td to base IMU^‡ ms	−1.612 ± 0.209 (10)	−1.680 ± 0.170 (10)	+0.068
distance to base IMU ‖p‖ mm	94.70 ± 0.19 (10)	93.58 ± 0.86 (10)	◆ +1.12
t_ga gyro−accel ms	0.081 ± 0.066 (10)	not in model

aux IMU – t265 IMU · 21 rows · ◆ 3

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	0.99736 ± 0.00139 (10)	0.99908 ± 0.00344 (10)	−0.00172
gyro Dw[1,1] scale y	1.00472 ± 0.00203 (10)	1.00911 ± 0.00276 (10)	◆ −0.00439
gyro Dw[2,2] scale z	1.00288 ± 0.00057 (10)	1.00142 ± 0.00208 (10)	+0.00146
gyro Dw[1,0] cross	0.00242 ± 0.00163 (10)	0.00617 ± 0.00408 (10)	−0.00376
gyro Dw[2,0] cross	0.00296 ± 0.00065 (10)	0.00080 ± 0.00264 (10)	+0.00216
gyro Dw[2,1] cross	−0.00170 ± 0.00141 (10)	−0.01204 ± 0.00351 (10)	◆ +0.01034
accel Da[0,0] scale x	1.01987 ± 0.00053 (10)	1.01601 ± 0.00611 (10)	+0.00386
accel Da[1,1] scale y	1.01927 ± 0.00160 (10)	1.01475 ± 0.00544 (10)	+0.00452
accel Da[2,2] scale z	1.01301 ± 0.00200 (10)	1.00913 ± 0.00206 (10)	◆ +0.00388
Tg[0,0] g-sensitivity mrad·s/m	0.006 ± 0.042 (10)	1.183 ± 0.753 (10)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.020 ± 0.067 (10)	5.084 ± 1.362 (10)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.018 ± 0.096 (10)	0.238 ± 0.726 (10)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	0.000 ± 0.040 (10)	−0.908 ± 0.691 (10)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.030 ± 0.050 (10)	0.003 ± 0.318 (10)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	−0.068 ± 0.080 (10)	1.923 ± 0.432 (10)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.043 ± 0.035 (10)	−1.374 ± 1.098 (10)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.031 ± 0.091 (10)	−6.191 ± 1.532 (10)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	0.026 ± 0.054 (10)	0.110 ± 0.803 (10)	≠ model
time delay td to base IMU^‡ ms	2.830 ± 0.212 (10)	2.753 ± 0.188 (10)	+0.077
distance to base IMU ‖p‖ mm	125.15 ± 0.19 (10)	124.41 ± 1.42 (10)	+0.74
t_ga gyro−accel ms	−6.564 ± 0.054 (10)	not in model

MVIS default IMU model (calyx) · 41 of 117 rows, model-independent only · ◆ 6

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – Blackfly
fx px	448.30 ± 0.61 (10)	448.30 ± 0.61 (10)	fixed 446.75 ± 2.17 (10)		0.00
fy px	447.62 ± 0.58 (10)	447.63 ± 0.58 (10)	fixed 444.89 ± 2.50 (10)		+0.02
cx px	370.78 ± 0.71 (10)	370.78 ± 0.72 (10)	fixed 368.67 ± 3.28 (10)		0.00
cy px	281.31 ± 0.69 (10)	281.31 ± 0.69 (10)	fixed 278.41 ± 1.44 (10)		0.00
dist k1	−0.31178 ± 0.00310 (10)	−0.31177 ± 0.00309 (10)	fixed −0.30796 ± 0.00996 (10)		+0.00001
dist k2	0.08088 ± 0.00268 (10)	0.08088 ± 0.00267 (10)	fixed 0.07731 ± 0.00889 (10)		0.00000
dist p1	−0.00033 ± 0.00010 (10)	−0.00033 ± 0.00010 (10)	fixed 0.00011 ± 0.00038 (10)		0.00000
dist p2	−0.00012 ± 0.00013 (10)	−0.00012 ± 0.00013 (10)	fixed 0.00014 ± 0.00095 (10)		0.00000
time offset td to base IMU^‡ ms	−4.293 ± 0.093 (10)	−4.279 ± 0.096 (10)	−5.438 ± 0.149 (10)	◆ +1.158	+0.014
distance to base IMU ‖p‖ mm	89.32 ± 1.41 (10)	89.35 ± 1.41 (10)	91.41 ± 3.64 (10)	−2.06	+0.03
reproj median px	0.349 ± 0.032 (10)	0.349 ± 0.032 (10)	0.483 ± 0.209 (10)	fit	fit
cam1 – t265 fisheye 1
fx px	287.08 ± 0.36 (10)	287.09 ± 0.36 (10)	fixed 285.72 ± 1.61 (10)		+0.01
fy px	286.70 ± 0.34 (10)	286.70 ± 0.34 (10)	fixed 285.13 ± 1.69 (10)		+0.01
cx px	415.32 ± 0.27 (10)	415.33 ± 0.27 (10)	fixed 415.83 ± 0.55 (10)		+0.01
cy px	403.53 ± 0.25 (10)	403.53 ± 0.26 (10)	fixed 404.09 ± 1.09 (10)		−0.01
dist k1	−0.00341 ± 0.00286 (10)	−0.00340 ± 0.00284 (10)	fixed 0.00165 ± 0.01882 (10)		+0.00001
dist k2	0.03197 ± 0.00592 (10)	0.03209 ± 0.00589 (10)	fixed 0.02307 ± 0.04201 (10)		+0.00012
dist k3	−0.03046 ± 0.00596 (10)	−0.03059 ± 0.00591 (10)	fixed −0.02549 ± 0.03728 (10)		−0.00013
dist k4	0.00456 ± 0.00208 (10)	0.00457 ± 0.00206 (10)	fixed 0.00321 ± 0.01100 (10)		+0.00001
baseline to Blackfly mm	71.267 ± 0.140 (10)	71.268 ± 0.139 (10)	fixed 71.856 ± 1.457 (10)		+0.001
time offset td to base IMU^‡ ms	−5.022 ± 0.149 (10)	−5.014 ± 0.155 (10)	−5.999 ± 0.247 (10)	◆ +0.984	+0.008
distance to base IMU ‖p‖ mm	130.10 ± 1.48 (10)	130.13 ± 1.48 (10)	133.92 ± 3.19 (10)	◆ −3.79	+0.03
reproj median px	0.225 ± 0.013 (10)	0.225 ± 0.013 (10)	0.272 ± 0.126 (10)	fit	fit
cam2 – t265 fisheye 2
fx px	287.02 ± 0.33 (10)	287.02 ± 0.33 (10)	fixed 285.37 ± 1.99 (10)		0.00
fy px	286.40 ± 0.32 (10)	286.41 ± 0.32 (10)	fixed 284.57 ± 2.08 (10)		+0.01
cx px	414.56 ± 0.32 (10)	414.57 ± 0.32 (10)	fixed 415.44 ± 0.59 (10)		0.00
cy px	397.96 ± 0.28 (10)	397.96 ± 0.28 (10)	fixed 398.62 ± 1.01 (10)		0.00
dist k1	−0.00397 ± 0.00264 (10)	−0.00397 ± 0.00264 (10)	fixed 0.00218 ± 0.03237 (10)		0.00000
dist k2	0.03443 ± 0.00611 (10)	0.03443 ± 0.00610 (10)	fixed 0.02374 ± 0.07849 (10)		0.00000
dist k3	−0.03139 ± 0.00675 (10)	−0.03146 ± 0.00673 (10)	fixed −0.02631 ± 0.07301 (10)		−0.00007
dist k4	0.00430 ± 0.00258 (10)	0.00431 ± 0.00257 (10)	fixed 0.00280 ± 0.02243 (10)		+0.00001
baseline to Blackfly mm	69.849 ± 0.327 (10)	69.849 ± 0.328 (10)	fixed 71.630 ± 2.455 (10)		0.000
time offset td to base IMU^‡ ms	−7.798 ± 0.237 (10)	−7.791 ± 0.237 (10)	−8.778 ± 0.244 (10)	◆ +0.987	+0.007
distance to base IMU ‖p‖ mm	111.39 ± 1.72 (10)	111.45 ± 1.70 (10)	114.03 ± 3.99 (10)	−2.58	+0.06
reproj median px	0.239 ± 0.014 (10)	0.239 ± 0.015 (10)	0.288 ± 0.140 (10)	fit	fit
aux IMU – GX3-25
time delay td to base IMU^‡ ms	−0.032 ± 0.209 (10)	−0.022 ± 0.208 (10)	−0.133 ± 0.189 (10)	+0.111	+0.011
distance to base IMU ‖p‖ mm	131.31 ± 0.20 (10)	131.31 ± 0.20 (10)	129.26 ± 0.88 (10)	◆ +2.05	0.00
aux IMU – GX3-35
time delay td to base IMU^‡ ms	−1.612 ± 0.209 (10)	−1.602 ± 0.209 (10)	−1.680 ± 0.170 (10)	+0.078	+0.010
distance to base IMU ‖p‖ mm	94.70 ± 0.19 (10)	94.70 ± 0.19 (10)	93.58 ± 0.86 (10)	◆ +1.12	0.00
aux IMU – t265 IMU
time delay td to base IMU^‡ ms	2.830 ± 0.212 (10)	2.832 ± 0.212 (10)	2.753 ± 0.188 (10)	+0.079	+0.002
distance to base IMU ‖p‖ mm	125.15 ± 0.19 (10)	125.16 ± 0.19 (10)	124.41 ± 1.42 (10)	+0.75	+0.01

VI-Rig: 4 IMUs + 4 cameras

Our rig with T265 and ELP cameras, global and rolling shutter. Dataset page.

VI-Rig 4 cams: which tool repeats more tightly, per family and group of parameters
Family	Focal + pp px	Distortion	Distance mm	Readout ms	Time offset ms	Cam–IMU distance mm	Gyro	Accel
T1 · camera n = 15 matched t265 GS fisheye stereo	MVIS ×1.6 ^a	MVIS ×1.6 ^ba	Not separated ^a	—	—	—	—	—
T2 · camera n = 13 matched ELP RS stereo + readout	not compared	not compared	not compared	MVIS ×4.2	—	—	—	—
T7 · cam+IMU n = 14 matched Mixed t265-f1 + ELP-L, aux Xsens	input, not compared	input, not compared	input, not compared	input, not compared	Tie ^cdefg	MVIS ×7.6 ^cdef	MVIS ×2.9K-IMU ^cdfh	MVIS ×4.4K-IMU ^cdfh
T8 · cam+IMU n = 3 matched ELP RS stereo + two IMUs (IMUs only)	MVIS only	MVIS only	MVIS only	MVIS only	n = 3 · too few ^cg	n = 3 · too few ^c	n = 3 · too few ^ch	n = 3 · too few ^ch

T1 · camera · t265 GS fisheye stereo More repeatable or tie: Focal + pp MVIS · Distortion MVIS · Distance not separated n = 15 matched · MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15

Focal + pp MVIS×1.6^a
Distortion MVIS×1.6^ba
DistanceNot separated^a

T2 · camera · ELP RS stereo + readout More repeatable or tie: Readout MVIS n = 13 matched · MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15

Readout MVIS×4.2
Focal + pp · Distortion · Distance: not compared

T7 · cam+IMU · Mixed t265-f1 + ELP-L, aux Xsens More repeatable or tie: Time offset tie · Cam–IMU distance MVIS · Gyro MVIS K-IMU · Accel MVIS K-IMU n = 14 matched · MVIS 15 of 15 · Kalibr 14 of 15 · Basalt not run

Time offsetTie^cdefg
Cam–IMU distance MVIS×7.6^cdef
Gyro MVIS×2.9K-IMU^cdfh
Accel MVIS×4.4K-IMU^cdfh
Focal + pp · Distortion · Distance · Readout: input, not compared

T8 · cam+IMU · ELP RS stereo + two IMUs (IMUs only) Too few recordings (n = 3) n = 3 matched · MVIS 15 of 15 · Kalibr 3 of 15 · Basalt not run

Time offsetToo few recordings (n = 3)^cg
Cam–IMU distanceToo few recordings (n = 3)^c
GyroToo few recordings (n = 3)^ch
AccelToo few recordings (n = 3)^ch
Focal + pp · Distortion · Distance · Readout: MVIS only

k = the looser tool's typical scatter divided by the tighter tool's (median over the group's parameters, on the recordings every compared tool ran). Click a cell for its numbers.

a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).
b Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
c Data settings still differ between the tools on this family.
d Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
e Kalibr ran ELP-L as a global-shutter camera with its stamps shifted by the MVIS td of each recording, so the ELP-L rows are left out here.
f Kalibr's t265 fisheye 1 intrinsics came from an older MVIS export.
g MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.
h MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Each cell is a separate comparison on the recordings every compared tool ran. Cells are not added up and nothing is ordered across datasets.

Not part of this comparison

T2 distortion coefficients of ELP-L (4 rows): other camera model: MVIS equidistant, Kalibr and Basalt radtan
T2 distortion coefficients of ELP-R (4 rows): other camera model: MVIS equidistant, Kalibr and Basalt radtan
T7 reprojection error (2 rows): fit quality, not repeatability and not accuracy
T7 IMU g-sensitivity Tg (18 rows): Kalibr absorbs it in its IMU model
T7 gyro–accel delay t_ga (2 rows): MVIS only
T8 focal length + principal point of ELP-L (4 rows): no Kalibr rolling-shutter run
T8 distortion coefficients of ELP-L (4 rows): no Kalibr rolling-shutter run
T8 rolling-shutter readout of ELP-L (1 row): no Kalibr rolling-shutter run
T8 camera–imu time offset of ELP-L (1 row): no Kalibr rolling-shutter run
T8 camera–imu distance of ELP-L (1 row): no Kalibr rolling-shutter run
T8 reproj median of ELP-L (1 row): no Kalibr rolling-shutter run
T8 focal length + principal point of ELP-R (4 rows): no Kalibr rolling-shutter run
T8 distortion coefficients of ELP-R (4 rows): no Kalibr rolling-shutter run
T8 rolling-shutter readout of ELP-R (1 row): no Kalibr rolling-shutter run
T8 camera–camera distance of ELP-R (1 row): no Kalibr rolling-shutter run
T8 camera–imu time offset of ELP-R (1 row): no Kalibr rolling-shutter run
T8 camera–imu distance of ELP-R (1 row): no Kalibr rolling-shutter run
T8 reproj median of ELP-R (1 row): no Kalibr rolling-shutter run
T8 IMU g-sensitivity Tg (18 rows): Kalibr absorbs it in its IMU model
T8 gyro–accel delay t_ga (2 rows): MVIS only
T7 Basalt: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig
T8 Basalt: not run: no rolling-shutter model

Focal length + principal point · T1 · 15 matched recordings (MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15) · t265 GS fisheye stereo, camera

Across the same 15 recordings, Kalibr's estimates scatter about 3.2× more than MVIS's (median over 8 parameters); Basalt's about 1.6×. A gap this large appeared in 0.049 % (Kalibr) and 0.0061 % (Basalt) of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 8 param.	0.31	0.15 – 0.5	0.00098 (raw 0.00049)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	0.62	0.36 – 0.83	0.00018 (raw 0.000061)	MVIS tighter	holds with any one recording left out
Kalibr ÷ Basalt · 8 param.	1.87	1.18 – 2.92	0.0055	Basalt tighter	holds with any one recording left out

The 95 % interval of MVIS ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

The 95 % interval of Kalibr ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

Parameter	MVIS sd	Kalibr sd	Basalt sd
t265 fisheye 1 · fx px	0.646	2.4	0.65
t265 fisheye 1 · fy px	0.722	2.59	0.791
t265 fisheye 1 · cx px	0.253	0.652	0.677
t265 fisheye 1 · cy px	0.336	0.849	0.799
t265 fisheye 2 · fx px	0.63	2.42	0.593
t265 fisheye 2 · fy px	0.707	2.57	0.746
t265 fisheye 2 · cx px	0.234	0.669	0.646
t265 fisheye 2 · cy px	0.329	0.839	0.784

a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T1 · 15 matched recordings (MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15) · t265 GS fisheye stereo, camera

Across the same 15 recordings, Kalibr's estimates scatter about 34× more than MVIS's (median over 8 parameters); Basalt's about 1.6×. A gap this large appeared in 0.0061 % (Kalibr) and 2.3 % (Basalt) of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 8 param.	0.03	0.01 – 0.11	0.00018 (raw 0.000061)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	0.61	0.4 – 0.91	0.023	MVIS tighter	holds with any one recording left out
Kalibr ÷ Basalt · 8 param.	19.97	5.54 – 38.03	0.00018 (raw 0.000061)	Basalt tighter	holds with any one recording left out

The 95 % interval of MVIS ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

Parameter	MVIS sd	Kalibr sd	Basalt sd
t265 fisheye 1 · dist k1	0.00199	0.0482	0.00332
t265 fisheye 1 · dist k2	0.00337	0.127	0.00651
t265 fisheye 1 · dist k3	0.00416	0.13	0.00637
t265 fisheye 1 · dist k4	0.00157	0.0441	0.00198
t265 fisheye 2 · dist k1	0.00211	0.0535	0.00325
t265 fisheye 2 · dist k2	0.00336	0.158	0.00889
t265 fisheye 2 · dist k3	0.00403	0.181	0.00823
t265 fisheye 2 · dist k4	0.00155	0.0691	0.0025

b Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T1 · 15 matched recordings (MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15) · t265 GS fisheye stereo, camera

MVIS is tighter than Kalibr; the other pairs are not separated. No single tool is marked.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 1 param.	0.34	0.23 – 0.53	0.0016 (raw 0.00055)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 1 param.	0.25	0.13 – 0.95	0.08 (raw 0.04)	tie: gap within chance
Kalibr ÷ Basalt · 1 param.	0.73	0.36 – 2.66	0.98	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
t265 fisheye 2 · baseline to t265 fisheye 1 mm	0.0338	0.0991	0.135

a Basalt's fisheye-2 frames are snapped onto fisheye-1 (1.8–5 ms apart).

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T2 · Not compared (MVIS 15 of 15 · Kalibr other model · Basalt shown only) · ELP RS stereo + readout, camera

other camera model: MVIS equidistant, Kalibr and Basalt radtan

Not in this comparison

Kalibr · other model · 8 rows: other camera model: MVIS equidistant, Kalibr and Basalt radtan
Basalt · shown only · 8 rows: global-shutter model of a rolling-shutter rig

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T2 · Not compared (MVIS only · Kalibr other model · Basalt other model) · ELP RS stereo + readout, camera

other camera model: MVIS equidistant, Kalibr and Basalt radtan

Not in this comparison

MVIS · only this tool · 8 rows: other camera model: MVIS equidistant, Kalibr and Basalt radtan
Kalibr · other model · 8 rows: other camera model: MVIS equidistant, Kalibr and Basalt radtan
Basalt · other model · 8 rows: other camera model: MVIS equidistant, Kalibr and Basalt radtan

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T2 · Not compared (MVIS – · Kalibr – · Basalt –) · ELP RS stereo + readout, camera

kept as text: Kalibr's rolling-shutter runs solve one camera per run

Not in this comparison

All tools · text note · 1 row: kept as text: Kalibr's rolling-shutter runs solve one camera per run

Repeatability only — a tool can repeat well and still be biased.

Rolling-shutter readout · T2 · 13 matched recordings (MVIS 15 of 15 · Kalibr 13–15 of 15 · Basalt shown only) · ELP RS stereo + readout, camera

Across the same 13 recordings, Kalibr's estimates scatter about 4.2× more than MVIS's (median over 2 parameters). A gap this large appeared in 0.073 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 2 param.	0.24	0.15 – 0.48	0.00073	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
ELP-L · readout tr ms	0.35 all: 0.409 (15)	1.49 all: 2 (15)	0.23
ELP-R · readout tr ms	0.346 all: 0.408 (15)	1.42	0.24

Not in this comparison

Basalt · shown only · 2 rows: global-shutter model of a rolling-shutter rig
MVIS · 2 recordings with values outside the matched set: another compared tool has no value there on some row
Kalibr · 2 recordings with values outside the matched set: another compared tool has no value there on some row

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T7 · Input, not compared (MVIS 14 of 15 · Kalibr input · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 8 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T7 · Input, not compared (MVIS 14 of 15 · Kalibr input · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 8 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T7 · Input, not compared (MVIS 14 of 15 · Kalibr input · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 1 row: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Rolling-shutter readout · T7 · Input, not compared (MVIS 14 of 15 · Kalibr input · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
Basalt · 1 row: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T7 · 14 matched recordings (MVIS 14 of 15 · Kalibr 14 of 15 · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

The spreads of MVIS and Kalibr differ by 1.1×, but a gap this size appears often by chance with 14 recordings (p = 0.35). Treated as a tie.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 2 param.	0.88	0.65 – 1.18	0.35	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Ratio
t265 fisheye 1 · time offset td to base IMU ms	0.188	0.301	0.62
ELP-L · time offset td to base IMU ms	0.64	–	–
Xsens MTi · time delay td to base IMU ms	0.152	0.122	1.25

Not in this comparison

Basalt · 3 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig
Kalibr · other model · 1 row: Kalibr treats ELP-L as global shutter, and its ELP-L stamps were shifted by the MVIS td of each recording before its run

c Data settings still differ between the tools on this family.
d Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
e Kalibr ran ELP-L as a global-shutter camera with its stamps shifted by the MVIS td of each recording, so the ELP-L rows are left out here.
f Kalibr's t265 fisheye 1 intrinsics came from an older MVIS export.
g MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T7 · 14 matched recordings (MVIS 14 of 15 · Kalibr 14 of 15 · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Across the same 14 recordings, Kalibr's estimates scatter about 7.6× more than MVIS's (median over 2 parameters). A gap this large appeared in 0.024 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 2 param.	0.13	0.09 – 0.24	0.00024	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
t265 fisheye 1 · distance to base IMU ‖p‖ mm	0.44	2.83	0.16
ELP-L · distance to base IMU ‖p‖ mm	0.447	–	–
Xsens MTi · distance to base IMU ‖p‖ mm	0.329	2.96	0.11

Not in this comparison

Basalt · 3 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig
Kalibr · other model · 1 row: Kalibr treats ELP-L as global shutter, and its ELP-L stamps were shifted by the MVIS td of each recording before its run

c Data settings still differ between the tools on this family.
d Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
e Kalibr ran ELP-L as a global-shutter camera with its stamps shifted by the MVIS td of each recording, so the ELP-L rows are left out here.
f Kalibr's t265 fisheye 1 intrinsics came from an older MVIS export.

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T7 · 14 matched recordings (MVIS 15 of 15 · Kalibr 14 of 15 · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Across the same 14 recordings, Kalibr's estimates scatter about 2.9× more than MVIS's (median over 12 parameters). A gap this large appeared in 0.049 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 12 param.	0.35	0.26 – 0.47	0.00049	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
t265 IMU · gyro Dw[0,0] scale x	0.00113 all: 0.00112 (15)	0.00365	0.31
t265 IMU · gyro Dw[1,1] scale y	0.00145 all: 0.00141 (15)	0.00268	0.54
t265 IMU · gyro Dw[2,2] scale z	0.000256 all: 0.000275 (15)	0.00049	0.52
t265 IMU · gyro Dw[1,0] cross	0.00132 all: 0.00132 (15)	0.0028	0.47
t265 IMU · gyro Dw[2,0] cross	0.00068 all: 0.000658 (15)	0.00327	0.21
t265 IMU · gyro Dw[2,1] cross	0.000481 all: 0.00054 (15)	0.00274	0.18
Xsens MTi · gyro Dw[0,0] scale x	0.00114 all: 0.0011 (15)	0.00356	0.32
Xsens MTi · gyro Dw[1,1] scale y	0.00021 all: 0.000206 (15)	0.000437	0.48
Xsens MTi · gyro Dw[2,2] scale z	0.00119 all: 0.00117 (15)	0.00274	0.43
Xsens MTi · gyro Dw[1,0] cross	0.000406 all: 0.000417 (15)	0.00283	0.14
Xsens MTi · gyro Dw[2,0] cross	0.000864 all: 0.000832 (15)	0.00232	0.37
Xsens MTi · gyro Dw[2,1] cross	0.000605 all: 0.000605 (15)	0.00264	0.23

Not in this comparison

Basalt · 12 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig
MVIS · 1 recording with values outside the matched set: another compared tool has no value there on some row

c Data settings still differ between the tools on this family.
d Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
f Kalibr's t265 fisheye 1 intrinsics came from an older MVIS export.
h MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T7 · 14 matched recordings (MVIS 15 of 15 · Kalibr 14 of 15 · Basalt not run) · Mixed t265-f1 + ELP-L, aux Xsens, cam+IMU

Across the same 14 recordings, Kalibr's estimates scatter about 4.4× more than MVIS's (median over 6 parameters). A gap this large appeared in 0.012 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 6 param.	0.23	0.15 – 0.3	0.00012	MVIS tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Ratio
t265 IMU · accel Da[0,0] scale x	0.00037 all: 0.000363 (15)	0.00143	0.26
t265 IMU · accel Da[1,1] scale y	0.000715 all: 0.000719 (15)	0.00177	0.40
t265 IMU · accel Da[2,2] scale z	0.000947 all: 0.00092 (15)	0.00221	0.43
Xsens MTi · accel Da[0,0] scale x	0.000336 all: 0.000324 (15)	0.0052	0.06
Xsens MTi · accel Da[1,1] scale y	0.000762 all: 0.000737 (15)	0.00478	0.16
Xsens MTi · accel Da[2,2] scale z	0.000628 all: 0.000675 (15)	0.00316	0.20

Not in this comparison

Basalt · 6 rows: not run: one IMU per run and no rolling-shutter model; no run holds the whole rig
MVIS · 1 recording with values outside the matched set: another compared tool has no value there on some row

c Data settings still differ between the tools on this family.
d Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
f Kalibr's t265 fisheye 1 intrinsics came from an older MVIS export.
h MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T8 · MVIS only (MVIS only · Kalibr no value · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

MVIS only: no Kalibr rolling-shutter run

Not in this comparison

MVIS · only this tool · 8 rows: no Kalibr rolling-shutter run
Kalibr · missing · 8 rows: no Kalibr rolling-shutter run
Basalt · 8 rows: not run: no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T8 · MVIS only (MVIS only · Kalibr no value · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

MVIS only: no Kalibr rolling-shutter run

Not in this comparison

MVIS · only this tool · 8 rows: no Kalibr rolling-shutter run
Kalibr · missing · 8 rows: no Kalibr rolling-shutter run
Basalt · 8 rows: not run: no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T8 · MVIS only (MVIS only · Kalibr no value · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

MVIS only: no Kalibr rolling-shutter run

Not in this comparison

MVIS · only this tool · 1 row: no Kalibr rolling-shutter run
Kalibr · missing · 1 row: no Kalibr rolling-shutter run
Basalt · 1 row: not run: no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Rolling-shutter readout · T8 · MVIS only (MVIS only · Kalibr no value · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

MVIS only: no Kalibr rolling-shutter run

Not in this comparison

MVIS · only this tool · 2 rows: no Kalibr rolling-shutter run
Kalibr · missing · 2 rows: no Kalibr rolling-shutter run
Basalt · 2 rows: not run: no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T8 · 3 matched recordings (MVIS 3 of 15 · Kalibr 3 of 15 · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

Only 3 recordings hold values from every compared tool. With 3 recordings the smallest possible p is 0.25, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 1 param.	1.39	not bounded at n = 3	0.5	too few recordings

Parameter	MVIS sd	Kalibr sd	Ratio
Xsens MTi · time delay td to base IMU ms	0.0378	0.0271	1.39

Not in this comparison

MVIS · only this tool · 2 rows: no Kalibr rolling-shutter run
Kalibr · missing · 2 rows: no Kalibr rolling-shutter run
Basalt · 3 rows: not run: no rolling-shutter model

c Data settings still differ between the tools on this family.
g MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T8 · 3 matched recordings (MVIS 3 of 15 · Kalibr 3 of 15 · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

Only 3 recordings hold values from every compared tool. With 3 recordings the smallest possible p is 0.25, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 1 param.	0.15	not bounded at n = 3	0.25	too few recordings

Parameter	MVIS sd	Kalibr sd	Ratio
Xsens MTi · distance to base IMU ‖p‖ mm	0.183	1.24	0.15

Not in this comparison

MVIS · only this tool · 2 rows: no Kalibr rolling-shutter run
Kalibr · missing · 2 rows: no Kalibr rolling-shutter run
Basalt · 3 rows: not run: no rolling-shutter model

c Data settings still differ between the tools on this family.

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T8 · 3 matched recordings (MVIS 15 of 15 · Kalibr 3 of 15 · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

Only 3 recordings hold values from every compared tool. With 3 recordings the smallest possible p is 0.25, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 12 param.	0.45	not bounded at n = 3	0.25	too few recordings

Parameter	MVIS sd	Kalibr sd	Ratio
t265 IMU · gyro Dw[0,0] scale x	0.000562 all: 0.00117 (15)	0.00388	0.14
t265 IMU · gyro Dw[1,1] scale y	0.00109 all: 0.00128 (15)	0.00259	0.42
t265 IMU · gyro Dw[2,2] scale z	0.000131 all: 0.000297 (15)	0.0000702	1.87
t265 IMU · gyro Dw[1,0] cross	0.00168 all: 0.00111 (15)	0.00141	1.19
t265 IMU · gyro Dw[2,0] cross	0.000535 all: 0.000735 (15)	0.00161	0.33
t265 IMU · gyro Dw[2,1] cross	0.000761 all: 0.00101 (15)	0.00312	0.24
Xsens MTi · gyro Dw[0,0] scale x	0.000668 all: 0.00119 (15)	0.00357	0.19
Xsens MTi · gyro Dw[1,1] scale y	0.000187 all: 0.000247 (15)	0.000132	1.42
Xsens MTi · gyro Dw[2,2] scale z	0.00108 all: 0.00131 (15)	0.00219	0.49
Xsens MTi · gyro Dw[1,0] cross	0.000497 all: 0.000541 (15)	0.000749	0.66
Xsens MTi · gyro Dw[2,0] cross	0.000827 all: 0.000693 (15)	0.0017	0.49
Xsens MTi · gyro Dw[2,1] cross	0.000465 all: 0.000785 (15)	0.00261	0.18

Not in this comparison

Basalt · 12 rows: not run: no rolling-shutter model
MVIS · 12 recordings with values outside the matched set: another compared tool has no value there on some row

c Data settings still differ between the tools on this family.
h MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T8 · 3 matched recordings (MVIS 15 of 15 · Kalibr 3 of 15 · Basalt not run) · ELP RS stereo + two IMUs (IMUs only), cam+IMU

Only 3 recordings hold values from every compared tool. With 3 recordings the smallest possible p is 0.25, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 6 param.	0.26	not bounded at n = 3	0.25	too few recordings

Parameter	MVIS sd	Kalibr sd	Ratio
t265 IMU · accel Da[0,0] scale x	0.00064 all: 0.000383 (15)	0.00204	0.31
t265 IMU · accel Da[1,1] scale y	0.000796 all: 0.000635 (15)	0.00206	0.39
t265 IMU · accel Da[2,2] scale z	0.000995 all: 0.00109 (15)	0.00141	0.71
Xsens MTi · accel Da[0,0] scale x	0.000543 all: 0.000366 (15)	0.00999	0.05
Xsens MTi · accel Da[1,1] scale y	0.000672 all: 0.000959 (15)	0.0068	0.10
Xsens MTi · accel Da[2,2] scale z	0.000813 all: 0.000492 (15)	0.00376	0.22

Not in this comparison

Basalt · 6 rows: not run: no rolling-shutter model
MVIS · 12 recordings with values outside the matched set: another compared tool has no value there on some row

c Data settings still differ between the tools on this family.
h MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Parameter tables, tool differences and run configs (VI-Rig 4 cams)

No ground truth: ◆ marks disagreement, not who is right.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

Key and MVIS column

Swipe a table sideways for the other tools and differences.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

T1 · t265 GS fisheye stereo

MVIS 15 of 15 · Kalibr 15 of 15 · Basalt 15 of 15

Camera · bags 01–15 · Kalibr pinhole-equi

Read first: MVIS estimates the board shape; Kalibr assumes a flat board.

Notes

Like for like: Bag 09: fisheye 2 starts one frame before fisheye 1. Kalibr ran on a copy with that frame dropped, because its 4 Hz sampling of each camera otherwise pairs frames one apart and it diverges.

Basalt: Basalt groups stereo frames by stamp and the t265 pair never shares one, so its fisheye 2 frames are snapped onto fisheye 1's (1.8–5 ms apart).

cam2 – t265 fisheye 1 · 8 rows · ◆ 6

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	286.54 ± 0.65 (15)	285.09 ± 2.40 (15)	284.14 ± 0.65 (15)	+1.45	−0.95	◆ −2.41
fy px	286.37 ± 0.72 (15)	284.18 ± 2.59 (15)	283.55 ± 0.79 (15)	+2.20	−0.63	◆ −2.82
cx px	415.52 ± 0.25 (15)	414.10 ± 0.65 (15)	413.73 ± 0.68 (15)	◆ +1.42	−0.37	◆ −1.79
cy px	404.33 ± 0.34 (15)	406.22 ± 0.85 (15)	407.13 ± 0.80 (15)	◆ −1.89	◆ +0.92	◆ +2.81
dist k1	−0.00779 ± 0.00199 (15)	−0.00487 ± 0.04819 (15)	−0.00863 ± 0.00332 (15)	−0.00292	−0.00376	−0.00084
dist k2	0.03606 ± 0.00337 (15)	0.04155 ± 0.12726 (15)	0.04117 ± 0.00651 (15)	−0.00549	−0.00038	+0.00512
dist k3	−0.03384 ± 0.00416 (15)	−0.04994 ± 0.12987 (15)	−0.04155 ± 0.00637 (15)	+0.01610	+0.00839	◆ −0.00771
dist k4	0.00564 ± 0.00157 (15)	0.01186 ± 0.04414 (15)	0.00859 ± 0.00198 (15)	−0.00622	−0.00327	◆ +0.00296

cam3 – t265 fisheye 2 · 9 rows · ◆ 5

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	286.54 ± 0.63 (15)	284.93 ± 2.42 (15)	284.02 ± 0.59 (15)	+1.61	−0.91	◆ −2.52
fy px	286.08 ± 0.71 (15)	284.48 ± 2.57 (15)	283.29 ± 0.75 (15)	+1.60	−1.19	◆ −2.79
cx px	415.06 ± 0.23 (15)	413.88 ± 0.67 (15)	413.82 ± 0.65 (15)	◆ +1.18	−0.06	◆ −1.24
cy px	398.64 ± 0.33 (15)	400.45 ± 0.84 (15)	401.47 ± 0.78 (15)	◆ −1.81	◆ +1.02	◆ +2.83
dist k1	−0.00903 ± 0.00211 (15)	−0.00739 ± 0.05345 (15)	−0.00705 ± 0.00325 (15)	−0.00164	+0.00034	+0.00198
dist k2	0.03775 ± 0.00336 (15)	0.04182 ± 0.15804 (15)	0.03298 ± 0.00889 (15)	−0.00407	−0.00884	−0.00477
dist k3	−0.03438 ± 0.00403 (15)	−0.04948 ± 0.18055 (15)	−0.03718 ± 0.00823 (15)	+0.01510	+0.01230	−0.00280
dist k4	0.00572 ± 0.00155 (15)	0.01283 ± 0.06911 (15)	0.00772 ± 0.00250 (15)	−0.00712	−0.00511	+0.00200
baseline to t265 fisheye 1 mm	63.663 ± 0.034 (15)	63.468 ± 0.099 (15)	63.555 ± 0.135 (15)	◆ +0.195	+0.087	−0.107

T2 · ELP RS stereo + readout

MVIS 15 of 15 · Kalibr 15 of 15 on ELP-L, 13 of 15 on ELP-R · Basalt 15 of 15, shown only

Camera · bags 01–15 · Kalibr pinhole-radtan-rs, one camera per run

Read first: Lens models differ, so only readout has a difference, and Kalibr wasn't re-run after the ELP stamp correction. Basalt is shown only.

Notes

Like for like: Kalibr has 13 of 15 recordings on ELP-R. MVIS uses an equidistant lens model, Kalibr and Basalt radtan, so fx, fy, cx, cy and distortion have no difference, and the dist[2] and dist[3] rows hold different coefficients per tool. Basalt solves this rolling-shutter pair as global shutter. Kalibr wasn't re-run after the ELP stamp correction; we expect it to be unchanged, but that wasn't tested. MVIS estimates the board shape; Kalibr assumes a flat board.

Note: Kalibr's rolling-shutter model calibrates one camera at a time, so it has no ELP stereo baseline. ELP-R bags 8 and 13 failed (autodiff overflow at bag-freq 15; the bag-freq-10 retry overflowed again on bag 8 and converged to a non-physical basin on bag 13) and are excluded.

Basalt: Basalt has no rolling-shutter model: it solves the ELP pair as global shutter with pinhole-radtan8 (seeded from kb4), so its T2 values are shown only. Since 2026-11-01 MVIS solves the ELP with the equidistant model while the Kalibr RS runs are radtan: readout is compared; fx, fy, cx, cy and distortion are shown without a difference because the lens models differ.

cam0 – ELP-L · 9 rows

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr
fx px	428.02 ± 0.77 (15)	427.70 ± 6.59 (15)	shown 427.07 ± 1.91 (15)	≠ model
fy px	427.21 ± 0.76 (15)	426.31 ± 6.42 (15)	shown 427.47 ± 2.23 (15)	≠ model
cx px	351.06 ± 0.46 (15)	338.22 ± 2.94 (15)	shown 348.63 ± 1.46 (15)	≠ model
cy px	267.80 ± 0.74 (15)	271.04 ± 8.33 (15)	shown 274.37 ± 1.48 (15)	≠ model
dist k1	−0.09126 ± 0.00172 (15)	other model −0.36695 ± 0.02361 (15)	shown 8.54347 ± 11.88550 (15)	≠ model
dist k2	−0.029 ± 0.006 (15)	other model 0.101 ± 0.020 (15)	shown 30.721 ± 18.766 (15)	≠ model
dist[2] (MVIS k3, radtan p1)	0.02039 ± 0.01227 (15)	other model −0.00041 ± 0.00285 (15)	shown −0.00195 ± 0.00054 (15)	≠ model
dist[3] (MVIS k4, radtan p2)	−0.00531 ± 0.00866 (15)	other model 0.00224 ± 0.00079 (15)	shown 0.00022 ± 0.00029 (15)	≠ model
readout tr ms	14.600 ± 0.409 (15)	16.354 ± 1.999 (15)	no RS model	−1.755

cam1 – ELP-R · 9 rows · ◆ 1

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr
fx px	424.79 ± 0.77 (15)	424.57 ± 3.89 (13/15)	shown 423.53 ± 1.86 (15)	≠ model
fy px	424.09 ± 0.76 (15)	424.57 ± 3.84 (13/15)	shown 423.70 ± 2.13 (15)	≠ model
cx px	336.85 ± 0.46 (15)	328.46 ± 2.34 (13/15)	shown 336.19 ± 1.47 (15)	≠ model
cy px	259.00 ± 0.74 (15)	261.01 ± 3.17 (13/15)	shown 266.06 ± 1.49 (15)	≠ model
dist k1	−0.08985 ± 0.00129 (15)	other model −0.35838 ± 0.00363 (13/15)	shown −3.68841 ± 8.37771 (15)	≠ model
dist k2	−0.025 ± 0.006 (15)	other model 0.094 ± 0.003 (13/15)	shown 26.167 ± 16.306 (15)	≠ model
dist[2] (MVIS k3, radtan p1)	0.00629 ± 0.01126 (15)	other model −0.00072 ± 0.00144 (13/15)	shown −0.00213 ± 0.00051 (15)	≠ model
dist[3] (MVIS k4, radtan p2)	0.00015 ± 0.00738 (15)	other model 0.00077 ± 0.00051 (13/15)	shown −0.00044 ± 0.00034 (15)	≠ model
readout tr ms	14.598 ± 0.408 (15)	17.369 ± 1.423 (13/15)	no RS model	◆ −2.771 ≠n

T7 · Mixed t265-f1 + ELP-L, aux Xsens

MVIS 15 of 15 · Kalibr 14 of 15 · Basalt not run

Camera + IMU · bags 01–15 · Kalibr with ELP-L as global shutter (bag 8 excluded)

Read first: Kalibr treats ELP-L as global shutter; its intrinsics and baselines are inputs from an MVIS export. Data settings differ: see Notes.

Notes, including why Basalt has no run

Basalt not run: Basalt cannot calibrate T7 — no rolling-shutter model for ELP-L, one IMU per run, and one time offset shared by all cameras (the t265 and ELP clocks share no frame stamps).

Like for like: Kalibr has 14 of 15 recordings (bag 8 failed). Its ELP-L readout is fixed at 0, and ELP-L's stamps were shifted by the MVIS td of the same bag before each Kalibr run. Its td and lever are solved given these inputs. MVIS uses two corners per tag, non-base cameras at 4 Hz or lower and an estimated board shape; Kalibr has none of these.

Note: MVIS solves t265 fisheye 1 (GS) + ELP-L (rolling shutter, readout estimated) + both IMUs. Kalibr's camera + IMU solver is global-shutter only, so its runs take ELP-L as a GS camera and its readout tr is reported as 0 (its global-shutter model, not an estimate). Kalibr bag 8 is excluded: its base gyro scales (1.23 / 2.23) are far from the other 14 bags (about 1.01), a failed solve; MVIS keeps bag 8, so its n is 15 and Kalibr's is 14.

cam0 – t265 fisheye 1 · 11 rows · ◆ 2

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
fx px	287.20 ± 0.43 (15)	fixed 287.20 ± 0.44 (14/15)
fy px	287.00 ± 0.52 (15)	fixed 287.09 ± 0.54 (14/15)
cx px	415.80 ± 0.35 (15)	fixed 415.78 ± 0.36 (14/15)
cy px	403.62 ± 0.28 (15)	fixed 403.62 ± 0.28 (14/15)
dist k1	−0.00588 ± 0.00125 (15)	fixed −0.00587 ± 0.00128 (14/15)
dist k2	0.03417 ± 0.00387 (15)	fixed 0.03411 ± 0.00393 (14/15)
dist k3	−0.03157 ± 0.00489 (15)	fixed −0.03143 ± 0.00498 (14/15)
dist k4	0.00457 ± 0.00197 (15)	fixed 0.00453 ± 0.00201 (14/15)
time offset td to base IMU^‡ ms	2.045 ± 0.182 (15)	−0.684 ± 0.301 (14/15)	◆ +2.729 ≠n
distance to base IMU ‖p‖ mm	11.92 ± 0.43 (15)	8.22 ± 2.83 (14/15)	◆ +3.69 ≠n
reproj median px	0.200 ± 0.005 (15)	0.367 ± 0.012 (14/15)	fit

cam1 – ELP-L · 13 rows

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
fx px	428.34 ± 0.71 (15)	fixed 428.22 ± 0.73 (14/15)
fy px	427.92 ± 0.86 (15)	fixed 427.89 ± 0.89 (14/15)
cx px	351.02 ± 0.35 (15)	fixed 351.01 ± 0.36 (14/15)
cy px	266.49 ± 0.25 (15)	fixed 266.47 ± 0.23 (14/15)
dist k1	−0.09405 ± 0.00194 (15)	fixed −0.09431 ± 0.00200 (14/15)
dist k2	−0.02494 ± 0.00707 (15)	fixed −0.02428 ± 0.00720 (14/15)
dist k3	0.02150 ± 0.01592 (15)	fixed 0.02013 ± 0.01638 (14/15)
dist k4	−0.01058 ± 0.01133 (15)	fixed −0.01005 ± 0.01172 (14/15)
readout tr ms	16.669 ± 0.086 (15)	fixed 0.000 (14/15)
baseline to t265 fisheye 1 mm	153.510 ± 0.164 (15)	fixed 153.504 ± 0.170 (14/15)
time offset td to base IMU^‡ ms	−115.919 ± 0.632 (15)	−107.879 ± 0.888 (14/15)	≠ model
distance to base IMU ‖p‖ mm	142.02 ± 0.43 (15)	146.46 ± 2.21 (14/15)	≠ model
reproj median px	0.181 ± 0.006 (15)	0.795 ± 0.031 (14/15)	fit

base IMU – t265 IMU · 19 rows · ◆ 5

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	1.00001 ± 0.00112 (15)	1.01343 ± 0.00365 (14/15)	◆ −0.01342 ≠n
gyro Dw[1,1] scale y	1.00321 ± 0.00141 (15)	1.01249 ± 0.00268 (14/15)	◆ −0.00928 ≠n
gyro Dw[2,2] scale z	1.00426 ± 0.00027 (15)	1.00560 ± 0.00049 (14/15)	◆ −0.00134 ≠n
gyro Dw[1,0] cross	0.00209 ± 0.00132 (15)	0.00378 ± 0.00280 (14/15)	−0.00169 ≠n
gyro Dw[2,0] cross	0.00257 ± 0.00066 (15)	−0.00599 ± 0.00327 (14/15)	◆ +0.00856 ≠n
gyro Dw[2,1] cross	0.00043 ± 0.00054 (15)	−0.00005 ± 0.00274 (14/15)	+0.00048 ≠n
accel Da[0,0] scale x	1.01998 ± 0.00036 (15)	1.01870 ± 0.00143 (14/15)	+0.00128 ≠n
accel Da[1,1] scale y	1.01901 ± 0.00072 (15)	1.01716 ± 0.00177 (14/15)	◆ +0.00185 ≠n
accel Da[2,2] scale z	1.01699 ± 0.00092 (15)	1.01709 ± 0.00221 (14/15)	−0.00010 ≠n
Tg[0,0] g-sensitivity mrad·s/m	−0.048 ± 0.106 (15)	0.040 ± 0.279 (14/15)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	−0.626 ± 0.121 (15)	2.415 ± 0.894 (14/15)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.009 ± 0.133 (15)	−0.844 ± 0.717 (14/15)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	0.193 ± 0.096 (15)	−1.014 ± 0.165 (14/15)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.088 ± 0.108 (15)	−1.293 ± 0.351 (14/15)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	−0.060 ± 0.143 (15)	2.639 ± 0.518 (14/15)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.011 ± 0.049 (15)	−0.101 ± 0.407 (14/15)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.268 ± 0.075 (15)	−2.886 ± 0.833 (14/15)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.094 ± 0.099 (15)	2.028 ± 0.593 (14/15)	≠ model
t_ga gyro−accel ms	−6.693 ± 0.108 (15)	not in model

aux IMU – Xsens MTi · 21 rows · ◆ 6

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	1.00096 ± 0.00110 (15)	1.01352 ± 0.00356 (14/15)	◆ −0.01256 ≠n
gyro Dw[1,1] scale y	0.99882 ± 0.00021 (15)	0.99985 ± 0.00044 (14/15)	◆ −0.00103 ≠n
gyro Dw[2,2] scale z	1.00027 ± 0.00117 (15)	1.00994 ± 0.00274 (14/15)	◆ −0.00967 ≠n
gyro Dw[1,0] cross	−0.00102 ± 0.00042 (15)	0.00559 ± 0.00283 (14/15)	◆ −0.00661 ≠n
gyro Dw[2,0] cross	0.00255 ± 0.00083 (15)	0.00404 ± 0.00232 (14/15)	−0.00148 ≠n
gyro Dw[2,1] cross	−0.00058 ± 0.00060 (15)	−0.00051 ± 0.00264 (14/15)	−0.00007 ≠n
accel Da[0,0] scale x	1.00005 ± 0.00032 (15)	1.00159 ± 0.00520 (14/15)	−0.00154 ≠n
accel Da[1,1] scale y	0.99984 ± 0.00074 (15)	1.01057 ± 0.00478 (14/15)	◆ −0.01073 ≠n
accel Da[2,2] scale z	1.00067 ± 0.00068 (15)	1.00287 ± 0.00316 (14/15)	−0.00220 ≠n
Tg[0,0] g-sensitivity mrad·s/m	0.013 ± 0.015 (15)	−0.121 ± 0.140 (14/15)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	−0.027 ± 0.075 (15)	0.832 ± 0.570 (14/15)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.033 ± 0.077 (15)	0.791 ± 0.462 (14/15)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.006 ± 0.020 (15)	−0.638 ± 0.306 (14/15)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	0.052 ± 0.058 (15)	0.708 ± 0.413 (14/15)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.059 ± 0.067 (15)	1.119 ± 0.429 (14/15)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.015 ± 0.021 (15)	−0.096 ± 0.150 (14/15)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	0.009 ± 0.076 (15)	−1.207 ± 0.294 (14/15)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.001 ± 0.043 (15)	−0.799 ± 0.254 (14/15)	≠ model
time delay td to base IMU^‡ ms	7.240 ± 0.146 (15)	7.260 ± 0.122 (14/15)	−0.019 ≠n
distance to base IMU ‖p‖ mm	124.09 ± 0.32 (15)	114.10 ± 2.96 (14/15)	◆ +9.99 ≠n
t_ga gyro−accel ms	−0.483 ± 0.218 (15)	not in model

MVIS default IMU model (calyx) · 26 of 64 rows, model-independent only · ◆ 3

These rows use Kalibr's recordings where Kalibr has them, so they compare the same recordings.

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – t265 fisheye 1
fx px	287.20 ± 0.43 (15)	287.20 ± 0.44 (14/15)	fixed 287.20 ± 0.44 (14/15)		0.00 ≠n
fy px	287.00 ± 0.52 (15)	287.09 ± 0.54 (14/15)	fixed 287.09 ± 0.54 (14/15)		+0.09 ≠n
cx px	415.80 ± 0.35 (15)	415.78 ± 0.36 (14/15)	fixed 415.78 ± 0.36 (14/15)		−0.02 ≠n
cy px	403.62 ± 0.28 (15)	403.62 ± 0.28 (14/15)	fixed 403.62 ± 0.28 (14/15)		−0.01 ≠n
dist k1	−0.00588 ± 0.00125 (15)	−0.00586 ± 0.00128 (14/15)	fixed −0.00587 ± 0.00128 (14/15)		+0.00002 ≠n
dist k2	0.03417 ± 0.00387 (15)	0.03410 ± 0.00394 (14/15)	fixed 0.03411 ± 0.00393 (14/15)		−0.00006 ≠n
dist k3	−0.03157 ± 0.00489 (15)	−0.03143 ± 0.00499 (14/15)	fixed −0.03143 ± 0.00498 (14/15)		+0.00014 ≠n
dist k4	0.00457 ± 0.00197 (15)	0.00454 ± 0.00201 (14/15)	fixed 0.00453 ± 0.00201 (14/15)		−0.00003 ≠n
time offset td to base IMU^‡ ms	2.045 ± 0.182 (15)	2.036 ± 0.188 (14/15)	−0.684 ± 0.301 (14/15)	◆ +2.720	−0.009 ≠n
distance to base IMU ‖p‖ mm	11.92 ± 0.43 (15)	11.99 ± 0.44 (14/15)	8.22 ± 2.83 (14/15)	◆ +3.76	+0.07 ≠n
reproj median px	0.200 ± 0.005 (15)	0.200 ± 0.005 (14/15)	0.367 ± 0.012 (14/15)	fit	fit
cam1 – ELP-L
fx px	428.34 ± 0.71 (15)	428.22 ± 0.74 (14/15)	fixed 428.22 ± 0.73 (14/15)		−0.12 ≠n
fy px	427.92 ± 0.86 (15)	427.89 ± 0.90 (14/15)	fixed 427.89 ± 0.89 (14/15)		−0.02 ≠n
cx px	351.02 ± 0.35 (15)	351.00 ± 0.36 (14/15)	fixed 351.01 ± 0.36 (14/15)		−0.02 ≠n
cy px	266.49 ± 0.25 (15)	266.47 ± 0.23 (14/15)	fixed 266.47 ± 0.23 (14/15)		−0.02 ≠n
dist k1	−0.09405 ± 0.00194 (15)	−0.09428 ± 0.00199 (14/15)	fixed −0.09431 ± 0.00200 (14/15)		−0.00023 ≠n
dist k2	−0.02494 ± 0.00707 (15)	−0.02435 ± 0.00720 (14/15)	fixed −0.02428 ± 0.00720 (14/15)		+0.00059 ≠n
dist k3	0.02150 ± 0.01592 (15)	0.02020 ± 0.01640 (14/15)	fixed 0.02013 ± 0.01638 (14/15)		−0.00130 ≠n
dist k4	−0.01058 ± 0.01133 (15)	−0.01000 ± 0.01174 (14/15)	fixed −0.01005 ± 0.01172 (14/15)		+0.00058 ≠n
readout tr ms	16.669 ± 0.086 (15)	16.670 ± 0.091 (14/15)	fixed 0.000 (14/15)		+0.001 ≠n
baseline to t265 fisheye 1 mm	153.510 ± 0.164 (15)	153.500 ± 0.173 (14/15)	fixed 153.504 ± 0.170 (14/15)		−0.010 ≠n
time offset td to base IMU^‡ ms	−115.919 ± 0.632 (15)	−115.876 ± 0.640 (14/15)	−107.879 ± 0.888 (14/15)	≠ model	+0.043 ≠n
distance to base IMU ‖p‖ mm	142.02 ± 0.43 (15)	141.98 ± 0.45 (14/15)	146.46 ± 2.21 (14/15)	≠ model	−0.04 ≠n
reproj median px	0.181 ± 0.006 (15)	0.181 ± 0.006 (14/15)	0.795 ± 0.031 (14/15)	fit	fit
aux IMU – Xsens MTi
time delay td to base IMU^‡ ms	7.240 ± 0.146 (15)	7.228 ± 0.152 (14/15)	7.260 ± 0.122 (14/15)	−0.032	−0.013 ≠n
distance to base IMU ‖p‖ mm	124.09 ± 0.32 (15)	124.10 ± 0.33 (14/15)	114.10 ± 2.96 (14/15)	◆ +9.99	+0.01 ≠n

T8 · ELP RS stereo + two IMUs (IMUs only)

MVIS 15 of 15 · Kalibr 3 of 15, no values on ELP-L, ELP-R · Basalt not run

Camera + IMU · MVIS bags 1–15 · Kalibr bags 1–3, t265 fisheye 1 only

Read first: Kalibr has 3 of 15 recordings, IMUs only, from earlier runs with an MVIS export as camera input. Data settings differ: see Notes.

Notes, including why Basalt has no run

Basalt not run: Basalt has no rolling-shutter model.

Like for like: Kalibr's values come from separate earlier runs of bags 1–3 with t265 fisheye 1 only, so MVIS − Kalibr compares 15 recordings with 3. MVIS uses two corners per tag, non-base cameras at 4 Hz or lower and an estimated board shape; Kalibr has none of these.

Note: No RS-anchored Kalibr run exists: Kalibr's camera + IMU solver cannot use the ELP rolling-shutter cameras. T8's IMUs (t265 base, Xsens aux) are the same physical IMUs on the same bags as T7, so they are compared with separate earlier Kalibr runs of bags 1–3 (t265 fisheye 1 only), not the T7 runs. Its ELP cameras have no Kalibr counterpart and are listed with MVIS values only.

cam0 – ELP-L · 12 rows

Kalibr: no run. Its camera + IMU solver can't use rolling-shutter cameras.

Parameter	MVISKalibr IMU model	Kalibr
fx px	428.05 ± 0.61 (15)	no run
fy px	427.90 ± 0.77 (15)	no run
cx px	350.94 ± 0.38 (15)	no run
cy px	266.09 ± 0.31 (15)	no run
dist k1	−0.09366 ± 0.00219 (15)	no run
dist k2	−0.02558 ± 0.00740 (15)	no run
dist k3	0.01529 ± 0.01502 (15)	no run
dist k4	−0.00738 ± 0.00997 (15)	no run
readout tr ms	16.620 ± 0.121 (15)	no run
time offset td to base IMU^‡ ms	−115.844 ± 0.617 (15)	no run
distance to base IMU ‖p‖ mm	143.01 ± 0.40 (15)	no run
reproj median px	0.162 ± 0.006 (15)	no run

cam1 – ELP-R · 13 rows

Kalibr: no run. Its camera + IMU solver can't use rolling-shutter cameras.

Parameter	MVISKalibr IMU model	Kalibr
fx px	424.86 ± 0.65 (15)	no run
fy px	424.71 ± 0.80 (15)	no run
cx px	336.80 ± 0.34 (15)	no run
cy px	257.25 ± 0.30 (15)	no run
dist k1	−0.09144 ± 0.00236 (15)	no run
dist k2	−0.02083 ± 0.00545 (15)	no run
dist k3	0.00062 ± 0.00939 (15)	no run
dist k4	0.00251 ± 0.00606 (15)	no run
readout tr ms	16.596 ± 0.109 (15)	no run
baseline to ELP-L mm	59.204 ± 0.085 (15)	no run
time offset td to base IMU^‡ ms	−115.879 ± 0.598 (15)	no run
distance to base IMU ‖p‖ mm	202.12 ± 0.42 (15)	no run
reproj median px	0.182 ± 0.007 (15)	no run

base IMU – t265 IMU · 19 rows · ◆ 7

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	0.99983 ± 0.00117 (15)	1.01272 ± 0.00388 (3/15)	◆ −0.01289 ≠n
gyro Dw[1,1] scale y	1.00294 ± 0.00128 (15)	1.01135 ± 0.00259 (3/15)	◆ −0.00841 ≠n
gyro Dw[2,2] scale z	1.00415 ± 0.00030 (15)	1.00439 ± 0.00007 (3/15)	−0.00024 ≠n
gyro Dw[1,0] cross	0.00185 ± 0.00111 (15)	0.00541 ± 0.00141 (3/15)	◆ −0.00356 ≠n
gyro Dw[2,0] cross	0.00221 ± 0.00073 (15)	−0.00330 ± 0.00161 (3/15)	◆ +0.00551 ≠n
gyro Dw[2,1] cross	−0.00101 ± 0.00101 (15)	0.00459 ± 0.00312 (3/15)	◆ −0.00559 ≠n
accel Da[0,0] scale x	1.01989 ± 0.00038 (15)	1.01725 ± 0.00204 (3/15)	◆ +0.00264 ≠n
accel Da[1,1] scale y	1.01905 ± 0.00063 (15)	1.01997 ± 0.00206 (3/15)	−0.00092 ≠n
accel Da[2,2] scale z	1.01668 ± 0.00109 (15)	1.02013 ± 0.00141 (3/15)	◆ −0.00345 ≠n
Tg[0,0] g-sensitivity mrad·s/m	−0.074 ± 0.093 (15)	0.068 ± 0.443 (3/15)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	−0.433 ± 0.260 (15)	1.885 ± 1.225 (3/15)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	0.211 ± 0.207 (15)	−1.393 ± 1.041 (3/15)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	0.253 ± 0.119 (15)	−0.803 ± 0.236 (3/15)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.074 ± 0.162 (15)	−0.968 ± 0.718 (3/15)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.243 ± 0.236 (15)	2.343 ± 0.104 (3/15)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.114 ± 0.064 (15)	0.679 ± 0.446 (3/15)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.089 ± 0.105 (15)	−2.238 ± 1.000 (3/15)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.034 ± 0.102 (15)	1.657 ± 0.502 (3/15)	≠ model
t_ga gyro−accel ms	−6.698 ± 0.135 (15)	not in model

aux IMU – Xsens MTi · 21 rows · ◆ 9

Parameter	MVISKalibr IMU model	Kalibr	MVIS− Kalibr
gyro Dw[0,0] scale x	1.00085 ± 0.00119 (15)	1.01189 ± 0.00357 (3/15)	◆ −0.01104 ≠n
gyro Dw[1,1] scale y	0.99862 ± 0.00025 (15)	0.99906 ± 0.00013 (3/15)	◆ −0.00044 ≠n
gyro Dw[2,2] scale z	1.00039 ± 0.00131 (15)	1.00753 ± 0.00219 (3/15)	◆ −0.00714 ≠n
gyro Dw[1,0] cross	−0.00010 ± 0.00054 (15)	0.00283 ± 0.00075 (3/15)	◆ −0.00294 ≠n
gyro Dw[2,0] cross	0.00255 ± 0.00069 (15)	0.00558 ± 0.00170 (3/15)	◆ −0.00303 ≠n
gyro Dw[2,1] cross	0.00097 ± 0.00079 (15)	−0.00485 ± 0.00261 (3/15)	◆ +0.00582 ≠n
accel Da[0,0] scale x	0.99997 ± 0.00037 (15)	0.98907 ± 0.00999 (3/15)	◆ +0.01090 ≠n
accel Da[1,1] scale y	0.99932 ± 0.00096 (15)	1.00747 ± 0.00680 (3/15)	◆ −0.00815 ≠n
accel Da[2,2] scale z	1.00093 ± 0.00049 (15)	1.00373 ± 0.00376 (3/15)	−0.00280 ≠n
Tg[0,0] g-sensitivity mrad·s/m	−0.008 ± 0.030 (15)	0.018 ± 0.121 (3/15)	≠ model
Tg[0,1] g-sensitivity mrad·s/m	−0.012 ± 0.092 (15)	1.221 ± 0.746 (3/15)	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.001 ± 0.050 (15)	0.722 ± 0.527 (3/15)	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.014 ± 0.021 (15)	−0.963 ± 0.154 (3/15)	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.012 ± 0.054 (15)	1.057 ± 0.271 (3/15)	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.004 ± 0.048 (15)	1.124 ± 0.514 (3/15)	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.024 ± 0.020 (15)	−0.012 ± 0.129 (3/15)	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.058 ± 0.078 (15)	−0.862 ± 0.330 (3/15)	≠ model
Tg[2,2] g-sensitivity mrad·s/m	−0.031 ± 0.036 (15)	−0.737 ± 0.296 (3/15)	≠ model
time delay td to base IMU^‡ ms	7.248 ± 0.115 (15)	7.254 ± 0.027 (3/15)	−0.006 ≠n
distance to base IMU ‖p‖ mm	123.94 ± 0.30 (15)	115.85 ± 1.24 (3/15)	◆ +8.09 ≠n
t_ga gyro−accel ms	−0.466 ± 0.237 (15)	not in model

MVIS default IMU model (calyx) · 27 of 65 rows, model-independent only · ◆ 2

Only the aux IMU td and lever rows are at Kalibr's 3 recordings.

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – ELP-L
fx px	428.05 ± 0.61 (15)	428.07 ± 0.62 (15)	no run		+0.02
fy px	427.90 ± 0.77 (15)	427.91 ± 0.77 (15)	no run		+0.01
cx px	350.94 ± 0.38 (15)	350.94 ± 0.38 (15)	no run		0.00
cy px	266.09 ± 0.31 (15)	266.08 ± 0.32 (15)	no run		0.00
dist k1	−0.09366 ± 0.00219 (15)	−0.09368 ± 0.00220 (15)	no run		−0.00003
dist k2	−0.02558 ± 0.00740 (15)	−0.02565 ± 0.00742 (15)	no run		−0.00008
dist k3	0.01529 ± 0.01502 (15)	0.01540 ± 0.01504 (15)	no run		+0.00012
dist k4	−0.00738 ± 0.00997 (15)	−0.00723 ± 0.00998 (15)	no run		+0.00015
readout tr ms	16.620 ± 0.121 (15)	16.618 ± 0.122 (15)	no run		−0.003
time offset td to base IMU^‡ ms	−115.844 ± 0.617 (15)	−115.884 ± 0.623 (15)	no run		−0.040
distance to base IMU ‖p‖ mm	143.01 ± 0.40 (15)	142.98 ± 0.39 (15)	no run		−0.03
reproj median px	0.162 ± 0.006 (15)	0.163 ± 0.006 (15)	no run		fit
cam1 – ELP-R
fx px	424.86 ± 0.65 (15)	424.87 ± 0.65 (15)	no run		+0.01
fy px	424.71 ± 0.80 (15)	424.73 ± 0.80 (15)	no run		+0.01
cx px	336.80 ± 0.34 (15)	336.80 ± 0.34 (15)	no run		0.00
cy px	257.25 ± 0.30 (15)	257.24 ± 0.30 (15)	no run		−0.01
dist k1	−0.09144 ± 0.00236 (15)	−0.09147 ± 0.00236 (15)	no run		−0.00002
dist k2	−0.02083 ± 0.00545 (15)	−0.02087 ± 0.00544 (15)	no run		−0.00004
dist k3	0.00062 ± 0.00939 (15)	0.00088 ± 0.00938 (15)	no run		+0.00026
dist k4	0.00251 ± 0.00606 (15)	0.00245 ± 0.00605 (15)	no run		−0.00006
readout tr ms	16.596 ± 0.109 (15)	16.594 ± 0.111 (15)	no run		−0.002
baseline to ELP-L mm	59.204 ± 0.085 (15)	59.207 ± 0.084 (15)	no run		+0.003
time offset td to base IMU^‡ ms	−115.879 ± 0.598 (15)	−115.931 ± 0.604 (15)	no run		−0.052
distance to base IMU ‖p‖ mm	202.12 ± 0.42 (15)	202.12 ± 0.41 (15)	no run		0.00
reproj median px	0.182 ± 0.007 (15)	0.183 ± 0.007 (15)	no run		fit
aux IMU – Xsens MTi
time delay td to base IMU^‡ ms	7.248 ± 0.115 (15)	7.176 ± 0.038 (3/15)	7.254 ± 0.027 (3/15)	◆ −0.078	−0.072 ≠n
distance to base IMU ‖p‖ mm	123.94 ± 0.30 (15)	124.19 ± 0.18 (3/15)	115.85 ± 1.24 (3/15)	◆ +8.34	+0.25 ≠n

TUM-VI

Calibration bags from TUM. Example.

TUM-VI: which tool repeats more tightly, per family and group of parameters
Family	Focal + pp px	Distortion	Distance mm	Time offset ms	Cam–IMU distance mm	Gyro	Accel
T3 · camera n = 8 matched Camera stereo (raw)	MVIS = Basalt> Kalibr ×1.5	Not separated ^a	Tie	—	—	—	—
T9 · cam+IMU n = 4 matched Camera + IMU, raw stream	n = 4 · too few	n = 4 · too few ^a	n = 4 · too few	n = 4 · too few ^b	n = 4 · too few	n = 4 · too few ^c	n = 4 · too few ^c
T10 · cam+IMU n = 4 matched Camera + IMU, standard-512	n = 4 · too few	n = 4 · too few ^a	n = 4 · too few	n = 4 · too few ^b	n = 4 · too few	n = 4 · too few ^c	n = 4 · too few ^c

T3 · camera · Camera stereo (raw) More repeatable or tie: Focal + pp MVIS = Basalt > Kalibr · Distortion not separated · Distance tie n = 8 matched · MVIS 8 of 8 · Kalibr 8 of 8 · Basalt 8 of 8

Focal + pp MVIS= Basalt > Kalibr×1.5
DistortionNot separated^a
DistanceTie

T9 · cam+IMU · Camera + IMU, raw stream Too few recordings (n = 4) n = 4 matched · MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Focal + ppToo few recordings (n = 4)
DistortionToo few recordings (n = 4)^a
DistanceToo few recordings (n = 4)
Time offsetToo few recordings (n = 4)^b
Cam–IMU distanceToo few recordings (n = 4)
GyroToo few recordings (n = 4)^c
AccelToo few recordings (n = 4)^c

T10 · cam+IMU · Camera + IMU, standard-512 Too few recordings (n = 4) n = 4 matched · MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Focal + ppToo few recordings (n = 4)
DistortionToo few recordings (n = 4)^a
DistanceToo few recordings (n = 4)
Time offsetToo few recordings (n = 4)^b
Cam–IMU distanceToo few recordings (n = 4)
GyroToo few recordings (n = 4)^c
AccelToo few recordings (n = 4)^c

k = the looser tool's typical scatter divided by the tighter tool's (median over the group's parameters, on the recordings every compared tool ran). Click a cell for its numbers.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
b MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.
c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Each cell is a separate comparison on the recordings every compared tool ran. Cells are not added up and nothing is ordered across datasets.

Not part of this comparison

T9 reprojection error (2 rows): fit quality, not repeatability and not accuracy
T9 IMU g-sensitivity Tg (9 rows): Kalibr absorbs it in its IMU model
T9 gyro–accel delay t_ga (1 row): MVIS only
T10 reprojection error (2 rows): fit quality, not repeatability and not accuracy
T10 IMU g-sensitivity Tg (9 rows): Kalibr absorbs it in its IMU model
T10 gyro–accel delay t_ga (1 row): MVIS only

Focal length + principal point · T3 · 8 matched recordings (MVIS 8 of 8 · Kalibr 8 of 8 · Basalt 8 of 8) · Camera stereo (raw), camera

The spreads of MVIS and Basalt differ by 1.1×, but a gap this size appears often by chance with 8 recordings (p = 0.44). Treated as a tie. Both scatter less than Kalibr: across the same 8 recordings Kalibr's estimates scatter about 1.5× more than MVIS's and 1.5× more than Basalt's. Holds with any one recording left out.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 8 param.	0.67	0.42 – 0.84	0.031 (raw 0.016)	MVIS tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	1.1	0.83 – 2.11	0.44	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	1.55	1.21 – 2.94	0.023 (raw 0.0078)	Basalt tighter	holds with any one recording left out

The 95 % interval of MVIS ÷ Kalibr reaches into the tie band (0.80–1.25): the ratio shown is a point value.

The 95 % interval of Kalibr ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · fx px	0.375	0.643	0.401
TUM-VI cam0 · fy px	0.409	0.656	0.411
TUM-VI cam0 · cx px	0.286	0.196	0.236
TUM-VI cam0 · cy px	0.165	0.236	0.11
TUM-VI cam1 · fx px	0.404	0.672	0.435
TUM-VI cam1 · fy px	0.426	0.666	0.428
TUM-VI cam1 · cx px	0.285	0.186	0.211
TUM-VI cam1 · cy px	0.379	0.324	0.306

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T3 · 8 matched recordings (MVIS 8 of 8 · Kalibr 8 of 8 · Basalt 8 of 8) · Camera stereo (raw), camera

Basalt is tighter than Kalibr; the other pairs are not separated. No single tool is marked.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 8 param.	0.42	0.13 – 0.95	0.078 (raw 0.039)	tie: gap within chance
MVIS ÷ Basalt · 8 param.	1.38	0.72 – 2.34	0.16	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	3.35	1.6 – 11.12	0.023 (raw 0.0078)	Basalt tighter	holds with any one recording left out

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · dist k1	0.00167	0.00392	0.00143
TUM-VI cam0 · dist k2	0.00284	0.00685	0.00203
TUM-VI cam0 · dist k3	0.00201	0.00488	0.00147
TUM-VI cam0 · dist k4	0.000498	0.00118	0.00039
TUM-VI cam1 · dist k1	0.00126	0.00318	0.00105
TUM-VI cam1 · dist k2	0.0022	0.00568	0.0015
TUM-VI cam1 · dist k3	0.0016	0.00384	0.000984
TUM-VI cam1 · dist k4	0.000389	0.000883	0.000242

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T3 · 8 matched recordings (MVIS 8 of 8 · Kalibr 8 of 8 · Basalt 8 of 8) · Camera stereo (raw), camera

The spreads of MVIS and Kalibr differ by 1.2×, but a gap this size appears often by chance with 8 recordings (p = 0.26). Treated as a tie. The spreads of MVIS and Basalt differ by 1.02×, but a gap this size appears often by chance with 8 recordings (p = 0.88). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.2×, but a gap this size appears often by chance with 8 recordings (p = 0.26). Treated as a tie.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 1 param.	0.82	0.27 – 1.06	0.26 (raw 0.086)	tie: gap within chance
MVIS ÷ Basalt · 1 param.	0.98	0.45 – 1.2	0.88	tie: gap within chance
Kalibr ÷ Basalt · 1 param.	1.2	0.93 – 1.83	0.26 (raw 0.11)	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam1 · baseline to TUM-VI cam0 mm	0.0804	0.0978	0.0818

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.84	not bounded at n = 4	0.75	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam0 · fx px	0.0861	0.111	0.78
TUM-VI cam0 · fy px	0.0842	0.119	0.71
TUM-VI cam0 · cx px	0.149	0.163	0.91
TUM-VI cam0 · cy px	0.0434	0.045	0.96
TUM-VI cam1 · fx px	0.0546	0.105	0.52
TUM-VI cam1 · fy px	0.0517	0.128	0.40
TUM-VI cam1 · cx px	0.128	0.135	0.95
TUM-VI cam1 · cy px	0.123	0.118	1.04

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.8	not bounded at n = 4	0.38	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam0 · dist k1	0.00107	0.00137	0.78
TUM-VI cam0 · dist k2	0.0019	0.00277	0.69
TUM-VI cam0 · dist k3	0.00126	0.00211	0.60
TUM-VI cam0 · dist k4	0.000271	0.00054	0.50
TUM-VI cam1 · dist k1	0.000731	0.000879	0.83
TUM-VI cam1 · dist k2	0.00133	0.00149	0.89
TUM-VI cam1 · dist k3	0.000988	0.00106	0.93
TUM-VI cam1 · dist k4	0.000259	0.000271	0.96

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 1 param.	1.35	not bounded at n = 4	0.75	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam1 · baseline to TUM-VI cam0 mm	0.00699	0.00519	1.35

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.08	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 2 param.	0.11	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 2 param.	1.48	not bounded at n = 4	0.38 (raw 0.25)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · time offset td to base IMU ms	0.0105	0.148	0.0999
TUM-VI cam1 · time offset td to base IMU ms	0.012	0.149	0.0999

b MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.96	not bounded at n = 4	1	too few recordings
MVIS ÷ Basalt · 2 param.	1.38	not bounded at n = 4	1 (raw 0.5)	too few recordings
Kalibr ÷ Basalt · 2 param.	1.43	not bounded at n = 4	1 (raw 0.62)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · distance to base IMU ‖p‖ mm	0.37	0.488	0.228
TUM-VI cam1 · distance to base IMU ‖p‖ mm	0.406	0.332	0.346

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 6 param.	0.57	not bounded at n = 4	0.75 (raw 0.25)	too few recordings
MVIS ÷ Basalt · 6 param.	0.92	not bounded at n = 4	0.75 (raw 0.5)	too few recordings
Kalibr ÷ Basalt · 6 param.	1.33	not bounded at n = 4	0.75 (raw 0.25)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
BMI160 imu0 · gyro Dw[0,0] scale x	0.000162	0.000416	0.000184
BMI160 imu0 · gyro Dw[1,1] scale y	0.000126	0.000135	0.000132
BMI160 imu0 · gyro Dw[2,2] scale z	0.000153	0.000222	0.000156
BMI160 imu0 · gyro Dw[1,0] cross	0.000166	0.000552	0.000443
BMI160 imu0 · gyro Dw[2,0] cross	0.000321	0.000682	0.000137
BMI160 imu0 · gyro Dw[2,1] cross	0.000123	0.000133	0.000447

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T9 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, raw stream, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 3 param.	3.07	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 3 param.	2.07	not bounded at n = 4	0.38	too few recordings
Kalibr ÷ Basalt · 3 param.	1.54	not bounded at n = 4	0.38 (raw 0.12)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
BMI160 imu0 · accel Da[0,0] scale x	0.000876	0.000269	0.0000544
BMI160 imu0 · accel Da[1,1] scale y	0.000439	0.000536	0.000348
BMI160 imu0 · accel Da[2,2] scale z	0.000905	0.000295	0.000437

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.84	not bounded at n = 4	0.62	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam0 · fx px	0.0849	0.18	0.47
TUM-VI cam0 · fy px	0.096	0.176	0.55
TUM-VI cam0 · cx px	0.072	0.0909	0.79
TUM-VI cam0 · cy px	0.0255	0.0265	0.96
TUM-VI cam1 · fx px	0.137	0.139	0.99
TUM-VI cam1 · fy px	0.157	0.13	1.21
TUM-VI cam1 · cx px	0.0558	0.0949	0.59
TUM-VI cam1 · cy px	0.0736	0.0833	0.88

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.33	not bounded at n = 4	0.25	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam0 · dist k1	0.000763	0.0023	0.33
TUM-VI cam0 · dist k2	0.00112	0.00488	0.23
TUM-VI cam0 · dist k3	0.0005	0.00394	0.13
TUM-VI cam0 · dist k4	0.000121	0.00102	0.12
TUM-VI cam1 · dist k1	0.00105	0.00171	0.61
TUM-VI cam1 · dist k2	0.00197	0.00424	0.46
TUM-VI cam1 · dist k3	0.00156	0.00415	0.38
TUM-VI cam1 · dist k4	0.000425	0.00129	0.33

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 1 param.	0.45	not bounded at n = 4	0.5	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
TUM-VI cam1 · baseline to TUM-VI cam0 mm	0.0243	0.0543	0.45

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, from one MVIS export used for all recordings, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.06	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 2 param.	0.08	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 2 param.	1.33	not bounded at n = 4	0.38 (raw 0.25)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · time offset td to base IMU ms	0.0116	0.165	0.124
TUM-VI cam1 · time offset td to base IMU ms	0.0084	0.165	0.124

b MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	1	not bounded at n = 4	1	too few recordings
MVIS ÷ Basalt · 2 param.	1.06	not bounded at n = 4	1	too few recordings
Kalibr ÷ Basalt · 2 param.	1.06	not bounded at n = 4	1 (raw 0.88)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
TUM-VI cam0 · distance to base IMU ‖p‖ mm	0.176	0.214	0.198
TUM-VI cam1 · distance to base IMU ‖p‖ mm	0.314	0.259	0.246

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 6 param.	0.38	not bounded at n = 4	0.75 (raw 0.25)	too few recordings
MVIS ÷ Basalt · 6 param.	0.6	not bounded at n = 4	0.75 (raw 0.25)	too few recordings
Kalibr ÷ Basalt · 6 param.	1.12	not bounded at n = 4	0.75	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
BMI160 imu0 · gyro Dw[0,0] scale x	0.000184	0.000349	0.000268
BMI160 imu0 · gyro Dw[1,1] scale y	0.000147	0.000159	0.000143
BMI160 imu0 · gyro Dw[2,2] scale z	0.000151	0.000333	0.000293
BMI160 imu0 · gyro Dw[1,0] cross	0.000119	0.00051	0.000489
BMI160 imu0 · gyro Dw[2,0] cross	0.000129	0.000808	0.00015
BMI160 imu0 · gyro Dw[2,1] cross	0.000108	0.000347	0.000559

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T10 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · Camera + IMU, standard-512, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 3 param.	0.61	not bounded at n = 4	0.5 (raw 0.25)	too few recordings
MVIS ÷ Basalt · 3 param.	0.52	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 3 param.	1.56	not bounded at n = 4	0.5 (raw 0.38)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
BMI160 imu0 · accel Da[0,0] scale x	0.0000368	0.000167	0.0000713
BMI160 imu0 · accel Da[1,1] scale y	0.00028	0.000462	0.000296
BMI160 imu0 · accel Da[2,2] scale z	0.000471	0.000466	0.000995

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Parameter tables, tool differences and run configs (TUM-VI)

No ground truth: ◆ marks disagreement, not who is right.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

Key and MVIS column

Swipe a table sideways for the other tools and differences.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

T3 · Camera stereo (raw)

MVIS 8 of 8 · Kalibr 8 of 8 · Basalt 8 of 8

Camera · cam1–cam8 · Kalibr kalibr_calibrate_cameras, pinhole-equi

Read first: Every tool ran on the same bags, filtered to view angles of 85° or less.

Notes

Note: Every tool ran on the same ≤ 85° view-filtered bags (TUM-VI corners beyond 85° break pinhole-equi in Kalibr).

cam0 – TUM-VI cam0 · 8 rows · ◆ 2

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	381.94 ± 0.37 (8)	381.77 ± 0.64 (8)	381.65 ± 0.40 (8)	+0.17	−0.12	−0.29
fy px	381.88 ± 0.41 (8)	381.73 ± 0.66 (8)	381.65 ± 0.41 (8)	+0.15	−0.08	−0.24
cx px	509.78 ± 0.29 (8)	510.06 ± 0.20 (8)	510.17 ± 0.24 (8)	−0.28	+0.11	◆ +0.38
cy px	513.81 ± 0.17 (8)	514.08 ± 0.24 (8)	514.11 ± 0.11 (8)	◆ −0.27	+0.03	◆ +0.30
dist k1	0.00500 ± 0.00167 (8)	0.00475 ± 0.00392 (8)	0.00523 ± 0.00143 (8)	+0.00025	+0.00048	+0.00023
dist k2	−0.00160 ± 0.00284 (8)	−0.00240 ± 0.00685 (8)	−0.00137 ± 0.00203 (8)	+0.00080	+0.00103	+0.00024
dist k3	−0.00048 ± 0.00201 (8)	0.00000 ± 0.00488 (8)	−0.00076 ± 0.00147 (8)	−0.00048	−0.00076	−0.00028
dist k4	−0.00011 ± 0.00050 (8)	−0.00023 ± 0.00118 (8)	−0.00004 ± 0.00039 (8)	+0.00012	+0.00019	+0.00008

cam1 – TUM-VI cam1 · 9 rows · ◆ 2

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	380.57 ± 0.40 (8)	380.70 ± 0.67 (8)	380.30 ± 0.43 (8)	−0.13	−0.40	−0.26
fy px	380.49 ± 0.43 (8)	380.65 ± 0.67 (8)	380.26 ± 0.43 (8)	−0.16	−0.38	−0.22
cx px	505.06 ± 0.28 (8)	505.35 ± 0.19 (8)	505.40 ± 0.21 (8)	◆ −0.30	+0.05	◆ +0.34
cy px	509.61 ± 0.38 (8)	509.94 ± 0.32 (8)	510.00 ± 0.31 (8)	−0.33	+0.05	◆ +0.38
dist k1	0.00662 ± 0.00126 (8)	0.00592 ± 0.00318 (8)	0.00660 ± 0.00105 (8)	+0.00070	+0.00067	−0.00003
dist k2	−0.00362 ± 0.00220 (8)	−0.00273 ± 0.00568 (8)	−0.00299 ± 0.00150 (8)	−0.00089	−0.00026	+0.00063
dist k3	0.00070 ± 0.00160 (8)	0.00039 ± 0.00384 (8)	0.00050 ± 0.00098 (8)	+0.00031	+0.00011	−0.00020
dist k4	−0.00035 ± 0.00039 (8)	−0.00036 ± 0.00088 (8)	−0.00036 ± 0.00024 (8)	+0.00001	0.00000	−0.00001
baseline to TUM-VI cam0 mm	101.048 ± 0.080 (8)	101.049 ± 0.098 (8)	101.047 ± 0.082 (8)	−0.001	−0.002	−0.001

T9 · Camera + IMU, raw stream

MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Camera + IMU · imu1–imu4 · Kalibr scale-misalignment vs MVIS (kalibr model)

Read first: Kalibr's intrinsics and baselines are inputs from an MVIS export; its td and lever are solved given them.

cam0 – TUM-VI cam0 · 11 rows · ◆ 4

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	382.71 ± 0.09 (4)	fixed 382.14 (4)	382.54 ± 0.11 (4)			◆ −0.17
fy px	382.67 ± 0.08 (4)	fixed 382.26 (4)	382.50 ± 0.12 (4)			◆ −0.17
cx px	509.90 ± 0.15 (4)	fixed 509.82 (4)	509.94 ± 0.16 (4)			+0.04
cy px	514.01 ± 0.04 (4)	fixed 514.51 (4)	514.09 ± 0.05 (4)			◆ +0.08
dist k1	0.00467 ± 0.00107 (4)	fixed 0.00477 (4)	0.00445 ± 0.00137 (4)			−0.00023
dist k2	−0.00326 ± 0.00190 (4)	fixed −0.00223 (4)	−0.00239 ± 0.00277 (4)			+0.00088
dist k3	0.00117 ± 0.00126 (4)	fixed −0.00007 (4)	0.00023 ± 0.00211 (4)			−0.00094
dist k4	−0.00057 ± 0.00027 (4)	fixed −0.00023 (4)	−0.00028 ± 0.00054 (4)			+0.00029
time offset td to base IMU^‡ ms	0.147 ± 0.014 (4)	−0.784 ± 0.148 (4)	−0.715 ± 0.100 (4)	◆ +0.931	+0.069	◆ −0.862
distance to base IMU ‖p‖ mm	94.96 ± 0.37 (4)	95.34 ± 0.49 (4)	95.01 ± 0.23 (4)	−0.38	−0.33	+0.05
reproj median px	0.096 ± 0.006 (4)	0.108 ± 0.007 (4)	not reported	fit

cam1 – TUM-VI cam1 · 12 rows · ◆ 4

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	381.29 ± 0.05 (4)	fixed 381.68 (4)	381.13 ± 0.11 (4)			◆ −0.17
fy px	381.27 ± 0.05 (4)	fixed 381.86 (4)	381.09 ± 0.13 (4)			◆ −0.18
cx px	505.30 ± 0.13 (4)	fixed 505.17 (4)	505.34 ± 0.13 (4)			+0.04
cy px	510.12 ± 0.12 (4)	fixed 510.40 (4)	510.09 ± 0.12 (4)			−0.02
dist k1	0.00604 ± 0.00073 (4)	fixed 0.00755 (4)	0.00589 ± 0.00088 (4)			−0.00014
dist k2	−0.00380 ± 0.00133 (4)	fixed −0.00667 (4)	−0.00298 ± 0.00149 (4)			+0.00082
dist k3	0.00130 ± 0.00099 (4)	fixed 0.00285 (4)	0.00050 ± 0.00106 (4)			−0.00080
dist k4	−0.00052 ± 0.00026 (4)	fixed −0.00087 (4)	−0.00031 ± 0.00027 (4)			+0.00021
baseline to TUM-VI cam0 mm	101.138 ± 0.007 (4)	fixed 101.004 ± 0.038 (4)	101.134 ± 0.005 (4)			−0.004
time offset td to base IMU^‡ ms	0.150 ± 0.015 (4)	−0.787 ± 0.149 (4)	−0.715 ± 0.100 (4)	◆ +0.937	+0.072	◆ −0.865
distance to base IMU ‖p‖ mm	100.04 ± 0.41 (4)	98.31 ± 0.33 (4)	99.62 ± 0.35 (4)	◆ +1.73	◆ +1.31	◆ −0.42
reproj median px	0.092 ± 0.005 (4)	0.109 ± 0.005 (4)	not reported	fit

base IMU – BMI160 imu0 · 19 rows · ◆ 7

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
gyro Dw[0,0] scale x	0.94568 ± 0.00016 (4)	0.94493 ± 0.00042 (4)	0.94562 ± 0.00018 (4)	◆ +0.00076	◆ +0.00070	−0.00006
gyro Dw[1,1] scale y	1.09414 ± 0.00013 (4)	1.09419 ± 0.00014 (4)	1.09416 ± 0.00013 (4)	−0.00005	−0.00003	+0.00002
gyro Dw[2,2] scale z	1.01744 ± 0.00015 (4)	1.01863 ± 0.00022 (4)	1.01740 ± 0.00016 (4)	◆ −0.00119	◆ −0.00123	−0.00004
gyro Dw[1,0] cross	0.00113 ± 0.00017 (4)	0.00202 ± 0.00055 (4)	0.00135 ± 0.00044 (4)	◆ −0.00088	◆ −0.00067	+0.00021
gyro Dw[2,0] cross	0.00052 ± 0.00032 (4)	0.00127 ± 0.00068 (4)	0.00058 ± 0.00014 (4)	◆ −0.00076	◆ −0.00069	+0.00006
gyro Dw[2,1] cross	0.00634 ± 0.00012 (4)	0.00497 ± 0.00013 (4)	0.00580 ± 0.00045 (4)	◆ +0.00137	◆ +0.00083	◆ −0.00054
accel Da[0,0] scale x	1.00717 ± 0.00088 (4)	1.00739 ± 0.00027 (4)	1.00736 ± 0.00005 (4)	−0.00022	−0.00003	+0.00019
accel Da[1,1] scale y	1.00696 ± 0.00044 (4)	1.00622 ± 0.00054 (4)	1.00665 ± 0.00035 (4)	◆ +0.00074	+0.00043	−0.00031
accel Da[2,2] scale z	0.97126 ± 0.00091 (4)	0.97202 ± 0.00030 (4)	0.97239 ± 0.00044 (4)	−0.00076	+0.00037	◆ +0.00114
Tg[0,0] g-sensitivity mrad·s/m	0.010 ± 0.013 (4)	0.016 ± 0.040 (4)	not in model	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.088 ± 0.032 (4)	1.043 ± 0.272 (4)	not in model	≠ model
Tg[0,2] g-sensitivity mrad·s/m	0.082 ± 0.080 (4)	0.152 ± 0.105 (4)	not in model	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.017 ± 0.030 (4)	−0.490 ± 0.324 (4)	not in model	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.031 ± 0.027 (4)	−0.049 ± 0.106 (4)	not in model	≠ model
Tg[1,2] g-sensitivity mrad·s/m	−0.026 ± 0.105 (4)	−0.015 ± 0.086 (4)	not in model	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.014 ± 0.027 (4)	−0.150 ± 0.041 (4)	not in model	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.075 ± 0.036 (4)	−0.092 ± 0.139 (4)	not in model	≠ model
Tg[2,2] g-sensitivity mrad·s/m	0.017 ± 0.020 (4)	0.010 ± 0.053 (4)	not in model	≠ model
t_ga gyro−accel ms	−1.508 ± 0.070 (4)	not in model	not in model

MVIS default IMU model (calyx) · 23 of 42 rows, model-independent only · ◆ 3

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – TUM-VI cam0
fx px	382.71 ± 0.09 (4)	382.71 ± 0.09 (4)	fixed 382.14 (4)		0.00
fy px	382.67 ± 0.08 (4)	382.67 ± 0.08 (4)	fixed 382.26 (4)		0.00
cx px	509.90 ± 0.15 (4)	509.90 ± 0.15 (4)	fixed 509.82 (4)		0.00
cy px	514.01 ± 0.04 (4)	514.01 ± 0.04 (4)	fixed 514.51 (4)		0.00
dist k1	0.00467 ± 0.00107 (4)	0.00467 ± 0.00107 (4)	fixed 0.00477 (4)		0.00000
dist k2	−0.00326 ± 0.00190 (4)	−0.00326 ± 0.00190 (4)	fixed −0.00223 (4)		0.00000
dist k3	0.00117 ± 0.00126 (4)	0.00117 ± 0.00126 (4)	fixed −0.00007 (4)		0.00000
dist k4	−0.00057 ± 0.00027 (4)	−0.00057 ± 0.00027 (4)	fixed −0.00023 (4)		0.00000
time offset td to base IMU^‡ ms	0.147 ± 0.014 (4)	0.172 ± 0.010 (4)	−0.784 ± 0.148 (4)	◆ +0.957	◆ +0.026
distance to base IMU ‖p‖ mm	94.96 ± 0.37 (4)	94.95 ± 0.37 (4)	95.34 ± 0.49 (4)	−0.39	0.00
reproj median px	0.096 ± 0.006 (4)	0.096 ± 0.006 (4)	0.108 ± 0.007 (4)	fit	fit
cam1 – TUM-VI cam1
fx px	381.29 ± 0.05 (4)	381.29 ± 0.05 (4)	fixed 381.68 (4)		0.00
fy px	381.27 ± 0.05 (4)	381.27 ± 0.05 (4)	fixed 381.86 (4)		0.00
cx px	505.30 ± 0.13 (4)	505.30 ± 0.13 (4)	fixed 505.17 (4)		0.00
cy px	510.12 ± 0.12 (4)	510.12 ± 0.12 (4)	fixed 510.40 (4)		0.00
dist k1	0.00604 ± 0.00073 (4)	0.00604 ± 0.00073 (4)	fixed 0.00755 (4)		0.00000
dist k2	−0.00380 ± 0.00133 (4)	−0.00380 ± 0.00133 (4)	fixed −0.00667 (4)		0.00000
dist k3	0.00130 ± 0.00099 (4)	0.00130 ± 0.00099 (4)	fixed 0.00285 (4)		0.00000
dist k4	−0.00052 ± 0.00026 (4)	−0.00052 ± 0.00026 (4)	fixed −0.00087 (4)		0.00000
baseline to TUM-VI cam0 mm	101.138 ± 0.007 (4)	101.138 ± 0.007 (4)	fixed 101.004 ± 0.038 (4)		0.000
time offset td to base IMU^‡ ms	0.150 ± 0.015 (4)	0.176 ± 0.012 (4)	−0.787 ± 0.149 (4)	◆ +0.962	◆ +0.025
distance to base IMU ‖p‖ mm	100.04 ± 0.41 (4)	100.04 ± 0.41 (4)	98.31 ± 0.33 (4)	◆ +1.74	+0.01
reproj median px	0.092 ± 0.005 (4)	0.092 ± 0.005 (4)	0.109 ± 0.005 (4)	fit	fit

T10 · Camera + IMU, standard-512

MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Camera + IMU · imu1–imu4 · Kalibr scale-misalignment vs MVIS (kalibr model)

Read first: Kalibr's intrinsics and baselines are inputs from an MVIS export; its td and lever are solved given them.

cam0 – TUM-VI cam0 · 11 rows · ◆ 6

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	191.51 ± 0.08 (4)	fixed 191.03 (4)	191.80 ± 0.18 (4)			◆ +0.29
fy px	191.49 ± 0.10 (4)	fixed 190.95 (4)	191.77 ± 0.18 (4)			◆ +0.29
cx px	254.59 ± 0.07 (4)	fixed 255.07 (4)	254.76 ± 0.09 (4)			◆ +0.16
cy px	256.65 ± 0.03 (4)	fixed 257.36 (4)	256.82 ± 0.03 (4)			◆ +0.18
dist k1	0.00367 ± 0.00076 (4)	fixed 0.00659 (4)	0.00259 ± 0.00230 (4)			−0.00108
dist k2	−0.00145 ± 0.00112 (4)	fixed −0.00617 (4)	−0.00049 ± 0.00488 (4)			+0.00096
dist k3	−0.00016 ± 0.00050 (4)	fixed 0.00315 (4)	−0.00056 ± 0.00394 (4)			−0.00040
dist k4	−0.00012 ± 0.00012 (4)	fixed −0.00106 (4)	−0.00014 ± 0.00102 (4)			−0.00002
time offset td to base IMU^‡ ms	0.182 ± 0.015 (4)	−0.723 ± 0.165 (4)	−0.644 ± 0.124 (4)	◆ +0.905	+0.079	◆ −0.827
distance to base IMU ‖p‖ mm	94.55 ± 0.18 (4)	95.60 ± 0.21 (4)	94.36 ± 0.20 (4)	◆ −1.05	◆ −1.24	−0.19
reproj median px	0.077 ± 0.002 (4)	0.081 ± 0.003 (4)	not reported	fit

cam1 – TUM-VI cam1 · 12 rows · ◆ 9

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	190.86 ± 0.14 (4)	fixed 190.50 (4)	191.12 ± 0.14 (4)			◆ +0.26
fy px	190.84 ± 0.16 (4)	fixed 190.48 (4)	191.11 ± 0.13 (4)			◆ +0.27
cx px	252.28 ± 0.06 (4)	fixed 252.85 (4)	252.44 ± 0.09 (4)			◆ +0.16
cy px	254.70 ± 0.07 (4)	fixed 255.10 (4)	254.84 ± 0.08 (4)			◆ +0.14
dist k1	0.00585 ± 0.00105 (4)	fixed 0.00791 (4)	0.00330 ± 0.00171 (4)			◆ −0.00255
dist k2	−0.00389 ± 0.00197 (4)	fixed −0.00655 (4)	0.00035 ± 0.00424 (4)			◆ +0.00424
dist k3	0.00149 ± 0.00156 (4)	fixed 0.00304 (4)	−0.00149 ± 0.00415 (4)			−0.00298
dist k4	−0.00061 ± 0.00042 (4)	fixed −0.00097 (4)	0.00017 ± 0.00129 (4)			+0.00078
baseline to TUM-VI cam0 mm	101.185 ± 0.024 (4)	fixed 101.120 ± 0.038 (4)	101.122 ± 0.054 (4)			◆ −0.063
time offset td to base IMU^‡ ms	0.182 ± 0.010 (4)	−0.730 ± 0.165 (4)	−0.644 ± 0.124 (4)	◆ +0.912	+0.086	◆ −0.827
distance to base IMU ‖p‖ mm	99.31 ± 0.31 (4)	99.66 ± 0.26 (4)	98.79 ± 0.25 (4)	◆ −0.36	◆ −0.88	◆ −0.52
reproj median px	0.078 ± 0.002 (4)	0.080 ± 0.003 (4)	not reported	fit

base IMU – BMI160 imu0 · 19 rows · ◆ 6

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
gyro Dw[0,0] scale x	1.00173 ± 0.00018 (4)	1.00244 ± 0.00035 (4)	1.00175 ± 0.00027 (4)	◆ −0.00071	◆ −0.00069	+0.00002
gyro Dw[1,1] scale y	0.99997 ± 0.00015 (4)	1.00015 ± 0.00016 (4)	1.00000 ± 0.00014 (4)	◆ −0.00018	−0.00015	+0.00003
gyro Dw[2,2] scale z	1.00117 ± 0.00015 (4)	1.00192 ± 0.00033 (4)	1.00087 ± 0.00029 (4)	◆ −0.00075	◆ −0.00105	◆ −0.00030
gyro Dw[1,0] cross	−0.00044 ± 0.00012 (4)	−0.00115 ± 0.00051 (4)	−0.00054 ± 0.00049 (4)	◆ +0.00072	◆ +0.00062	−0.00010
gyro Dw[2,0] cross	0.00092 ± 0.00013 (4)	0.00079 ± 0.00081 (4)	0.00088 ± 0.00015 (4)	+0.00014	+0.00010	−0.00004
gyro Dw[2,1] cross	0.00086 ± 0.00011 (4)	−0.00061 ± 0.00035 (4)	0.00042 ± 0.00056 (4)	◆ +0.00147	◆ +0.00103	−0.00044
accel Da[0,0] scale x	1.00328 ± 0.00004 (4)	1.00296 ± 0.00017 (4)	1.00293 ± 0.00007 (4)	◆ +0.00032	−0.00003	◆ −0.00035
accel Da[1,1] scale y	1.00527 ± 0.00028 (4)	1.00515 ± 0.00046 (4)	1.00527 ± 0.00030 (4)	+0.00012	+0.00012	0.00000
accel Da[2,2] scale z	1.00203 ± 0.00047 (4)	1.00194 ± 0.00047 (4)	1.00148 ± 0.00099 (4)	+0.00009	−0.00046	−0.00055
Tg[0,0] g-sensitivity mrad·s/m	0.015 ± 0.011 (4)	0.020 ± 0.045 (4)	not in model	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.001 ± 0.009 (4)	0.964 ± 0.274 (4)	not in model	≠ model
Tg[0,2] g-sensitivity mrad·s/m	0.006 ± 0.014 (4)	0.108 ± 0.102 (4)	not in model	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.015 ± 0.016 (4)	−0.522 ± 0.343 (4)	not in model	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.015 ± 0.022 (4)	−0.032 ± 0.093 (4)	not in model	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.010 ± 0.013 (4)	−0.009 ± 0.103 (4)	not in model	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.005 ± 0.056 (4)	−0.096 ± 0.020 (4)	not in model	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.027 ± 0.033 (4)	−0.076 ± 0.098 (4)	not in model	≠ model
Tg[2,2] g-sensitivity mrad·s/m	0.009 ± 0.025 (4)	0.002 ± 0.035 (4)	not in model	≠ model
t_ga gyro−accel ms	−1.641 ± 0.057 (4)	not in model	not in model

MVIS default IMU model (calyx) · 23 of 42 rows, model-independent only · ◆ 4

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – TUM-VI cam0
fx px	191.51 ± 0.08 (4)	191.51 ± 0.08 (4)	fixed 191.03 (4)		0.00
fy px	191.49 ± 0.10 (4)	191.49 ± 0.10 (4)	fixed 190.95 (4)		0.00
cx px	254.59 ± 0.07 (4)	254.59 ± 0.07 (4)	fixed 255.07 (4)		0.00
cy px	256.65 ± 0.03 (4)	256.65 ± 0.03 (4)	fixed 257.36 (4)		0.00
dist k1	0.00367 ± 0.00076 (4)	0.00367 ± 0.00076 (4)	fixed 0.00659 (4)		0.00000
dist k2	−0.00145 ± 0.00112 (4)	−0.00145 ± 0.00112 (4)	fixed −0.00617 (4)		0.00000
dist k3	−0.00016 ± 0.00050 (4)	−0.00016 ± 0.00050 (4)	fixed 0.00315 (4)		0.00000
dist k4	−0.00012 ± 0.00012 (4)	−0.00012 ± 0.00012 (4)	fixed −0.00106 (4)		0.00000
time offset td to base IMU^‡ ms	0.182 ± 0.015 (4)	0.184 ± 0.012 (4)	−0.723 ± 0.165 (4)	◆ +0.907	+0.002
distance to base IMU ‖p‖ mm	94.55 ± 0.18 (4)	94.55 ± 0.18 (4)	95.60 ± 0.21 (4)	◆ −1.05	0.00
reproj median px	0.077 ± 0.002 (4)	0.077 ± 0.002 (4)	0.081 ± 0.003 (4)	fit	fit
cam1 – TUM-VI cam1
fx px	190.86 ± 0.14 (4)	190.86 ± 0.14 (4)	fixed 190.50 (4)		0.00
fy px	190.84 ± 0.16 (4)	190.84 ± 0.16 (4)	fixed 190.48 (4)		0.00
cx px	252.28 ± 0.06 (4)	252.28 ± 0.06 (4)	fixed 252.85 (4)		0.00
cy px	254.70 ± 0.07 (4)	254.70 ± 0.07 (4)	fixed 255.10 (4)		0.00
dist k1	0.00585 ± 0.00105 (4)	0.00585 ± 0.00105 (4)	fixed 0.00791 (4)		0.00000
dist k2	−0.00389 ± 0.00197 (4)	−0.00389 ± 0.00197 (4)	fixed −0.00655 (4)		0.00000
dist k3	0.00149 ± 0.00156 (4)	0.00149 ± 0.00156 (4)	fixed 0.00304 (4)		0.00000
dist k4	−0.00061 ± 0.00042 (4)	−0.00061 ± 0.00042 (4)	fixed −0.00097 (4)		0.00000
baseline to TUM-VI cam0 mm	101.185 ± 0.024 (4)	101.185 ± 0.024 (4)	fixed 101.120 ± 0.038 (4)		0.000
time offset td to base IMU^‡ ms	0.182 ± 0.010 (4)	0.184 ± 0.008 (4)	−0.730 ± 0.165 (4)	◆ +0.914	+0.002
distance to base IMU ‖p‖ mm	99.31 ± 0.31 (4)	99.31 ± 0.31 (4)	99.66 ± 0.26 (4)	◆ −0.36	0.00
reproj median px	0.078 ± 0.002 (4)	0.078 ± 0.002 (4)	0.080 ± 0.003 (4)	fit	fit

Looper Insight-9

Two global-shutter cameras, one rolling-shutter camera and a MEMS IMU, from Looper Robotics. Example.

Looper Insight-9: which tool repeats more tightly, per family and group of parameters
Family	Focal + pp px	Distortion	Distance mm	Readout ms	Time offset ms	Cam–IMU distance mm	Gyro	Accel
T4 · camera n = 10 matched Camera stereo	Kalibr ×1.4	Tie ^a	Tie	—	—	—	—	—
T11 · cam+IMU n = 10 matched 3-cam + IMU	MVIS ×1.8	Tie ^a	Tie	MVIS only	Not separated ^b	Tie	Tie ^c	Tie ^c

T4 · camera · Camera stereo More repeatable or tie: Focal + pp Kalibr · Distortion tie · Distance tie n = 10 matched · MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10

Focal + pp Kalibr×1.4
DistortionTie^a
DistanceTie

T11 · cam+IMU · 3-cam + IMU More repeatable or tie: Focal + pp MVIS · Distortion tie · Distance tie · Time offset not separated · Cam–IMU distance tie · Gyro tie · Accel tie n = 10 matched · MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10

Focal + pp MVIS×1.8
DistortionTie^a
DistanceTie
Time offsetNot separated^b
Cam–IMU distanceTie
GyroTie^c
AccelTie^c
Readout: MVIS only

k = the looser tool's typical scatter divided by the tighter tool's (median over the group's parameters, on the recordings every compared tool ran). Click a cell for its numbers.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
b MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.
c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Each cell is a separate comparison on the recordings every compared tool ran. Cells are not added up and nothing is ordered across datasets.

Not part of this comparison

T11 reprojection error (2 rows): fit quality, not repeatability and not accuracy
T11 focal length + principal point of Looper RS cam2 (4 rows): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 distortion coefficients of Looper RS cam2 (4 rows): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 rolling-shutter readout of Looper RS cam2 (1 row): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 camera–camera distance of Looper RS cam2 (1 row): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 camera–imu time offset of Looper RS cam2 (1 row): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 camera–imu distance of Looper RS cam2 (1 row): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 reproj median of Looper RS cam2 (1 row): no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
T11 IMU g-sensitivity Tg (9 rows): Kalibr absorbs it in its IMU model
T11 gyro–accel delay t_ga (1 row): MVIS only

Focal length + principal point · T4 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · Camera stereo, camera

Across the same 10 recordings, MVIS's estimates scatter about 1.7× more than Kalibr's (median over 8 parameters); Basalt's about 1.4×. A gap this large appeared in 0.59 % (MVIS) and 1.6 % (Basalt) of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 8 param.	1.68	1.24 – 2.3	0.018 (raw 0.0059)	Kalibr tighter	holds with any one recording left out
MVIS ÷ Basalt · 8 param.	1.04	0.85 – 1.22	0.56	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	0.71	0.58 – 0.92	0.031 (raw 0.016)	Kalibr tighter	holds with any one recording left out

The 95 % interval of MVIS ÷ Kalibr reaches into the tie band (0.80–1.25): the ratio shown is a point value.

The 95 % interval of Kalibr ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper GS cam0 · fx px	0.379	0.187	0.383
Looper GS cam0 · fy px	0.374	0.186	0.38
Looper GS cam0 · cx px	0.106	0.169	0.109
Looper GS cam0 · cy px	0.0451	0.0548	0.0321
Looper GS cam1 · fx px	0.398	0.202	0.382
Looper GS cam1 · fy px	0.399	0.197	0.381
Looper GS cam1 · cx px	0.117	0.162	0.0915
Looper GS cam1 · cy px	0.0701	0.0489	0.0516

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T4 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · Camera stereo, camera

The spreads of MVIS and Kalibr differ by 1.6×, but a gap this size appears often by chance with 10 recordings (p = 0.21). Treated as a tie. The spreads of MVIS and Basalt differ by 1.2×, but a gap this size appears often by chance with 10 recordings (p = 0.21). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.8×, but a gap this size appears often by chance with 10 recordings (p = 0.21). Treated as a tie.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 8 param.	0.62	0.39 – 1.13	0.21 (raw 0.068)	tie: gap within chance
MVIS ÷ Basalt · 8 param.	1.22	0.94 – 1.52	0.21 (raw 0.11)	tie: gap within chance
Kalibr ÷ Basalt · 8 param.	1.77	0.96 – 2.62	0.21 (raw 0.08)	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper GS cam0 · dist k1	0.00109	0.00139	0.000789
Looper GS cam0 · dist k2	0.00284	0.00481	0.00156
Looper GS cam0 · dist k3	0.00381	0.00588	0.00177
Looper GS cam0 · dist k4	0.0017	0.00244	0.000639
Looper GS cam1 · dist k1	0.0011	0.00095	0.00103
Looper GS cam1 · dist k2	0.00224	0.00393	0.00226
Looper GS cam1 · dist k3	0.00259	0.00515	0.00288
Looper GS cam1 · dist k4	0.00103	0.00219	0.00126

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T4 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · Camera stereo, camera

The spreads of MVIS and Kalibr differ by 1.4×, but a gap this size appears often by chance with 10 recordings (p = 0.72). Treated as a tie. The spreads of MVIS and Basalt differ by 1.04×, but a gap this size appears often by chance with 10 recordings (p = 0.79). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.3×, but a gap this size appears often by chance with 10 recordings (p = 0.72). Treated as a tie.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 1 param.	0.74	0.56 – 1.31	0.72 (raw 0.24)	tie: gap within chance
MVIS ÷ Basalt · 1 param.	0.96	0.72 – 1.57	0.79	tie: gap within chance
Kalibr ÷ Basalt · 1 param.	1.3	0.8 – 1.72	0.72 (raw 0.32)	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper GS cam1 · baseline to Looper GS cam0 mm	0.176	0.238	0.183

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr input · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

Across the same 10 recordings, Basalt's estimates scatter about 1.8× more than MVIS's (median over 8 parameters). A gap this large appeared in 4.9 % of recording-by-recording swaps. Holds with any one recording left out.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.57	0.34 – 0.98	0.049	MVIS tighter	holds with any one recording left out

The 95 % interval of MVIS ÷ Basalt reaches into the tie band (0.80–1.25): the ratio shown is a point value.

Parameter	MVIS sd	Basalt sd	Ratio
Looper GS cam0 · fx px	0.155	0.44	0.35
Looper GS cam0 · fy px	0.164	0.435	0.38
Looper GS cam0 · cx px	0.132	0.154	0.86
Looper GS cam0 · cy px	0.065	0.0408	1.59
Looper GS cam1 · fx px	0.121	0.449	0.27
Looper GS cam1 · fy px	0.13	0.443	0.29
Looper GS cam1 · cx px	0.119	0.0931	1.28
Looper GS cam1 · cy px	0.0818	0.0562	1.46

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
MVIS · only this tool · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr input · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

The spreads of MVIS and Basalt differ by 1.3×, but a gap this size appears often by chance with 10 recordings (p = 0.17). Treated as a tie.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	1.3	0.86 – 1.75	0.17	tie: gap within chance

Parameter	MVIS sd	Basalt sd	Ratio
Looper GS cam0 · dist k1	0.00123	0.0012	1.03
Looper GS cam0 · dist k2	0.00408	0.00243	1.68
Looper GS cam0 · dist k3	0.00483	0.00284	1.70
Looper GS cam0 · dist k4	0.00187	0.00111	1.68
Looper GS cam1 · dist k1	0.000887	0.00137	0.65
Looper GS cam1 · dist k2	0.00328	0.00265	1.24
Looper GS cam1 · dist k3	0.00426	0.00328	1.30
Looper GS cam1 · dist k4	0.00174	0.00133	1.31

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
MVIS · only this tool · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 4 rows: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr input · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

The spreads of MVIS and Basalt differ by 1.1×, but a gap this size appears often by chance with 10 recordings (p = 0.38). Treated as a tie.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 1 param.	1.11	0.9 – 1.39	0.38	tie: gap within chance

Parameter	MVIS sd	Basalt sd	Ratio
Looper GS cam1 · baseline to Looper GS cam0 mm	0.211	0.191	1.10

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.
MVIS · only this tool · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Rolling-shutter readout · T11 · MVIS only (MVIS only · Kalibr no value · Basalt no value) · 3-cam + IMU, cam+IMU

MVIS only: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

Not in this comparison

MVIS · only this tool · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

MVIS is tighter than Basalt; the other pairs are not separated. No single tool is marked.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Kalibr · 2 param.	0.59	0.36 – 1.19	0.22 (raw 0.11)	tie: gap within chance
MVIS ÷ Basalt · 2 param.	0.5	0.34 – 0.65	0.012 (raw 0.0039)	MVIS tighter	holds with any one recording left out
Kalibr ÷ Basalt · 2 param.	0.85	0.43 – 1.28	0.43	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper GS cam0 · time offset td to base IMU ms	0.0767	0.131	0.156
Looper GS cam1 · time offset td to base IMU ms	0.0801	0.135	0.156

Not in this comparison

MVIS · only this tool · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

b MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

The spreads of MVIS and Kalibr differ by 1.1×, but a gap this size appears often by chance with 10 recordings (p = 1). Treated as a tie. The spreads of MVIS and Basalt differ by 1.2×, but a gap this size appears often by chance with 10 recordings (p = 1). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.03×, but a gap this size appears often by chance with 10 recordings (p = 1). Treated as a tie.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.9	0.67 – 1.24	1 (raw 0.39)	tie: gap within chance
MVIS ÷ Basalt · 2 param.	0.87	0.56 – 1.28	1 (raw 0.47)	tie: gap within chance
Kalibr ÷ Basalt · 2 param.	0.97	0.59 – 1.57	1 (raw 0.87)	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper GS cam0 · distance to base IMU ‖p‖ mm	0.375	0.512	0.564
Looper GS cam1 · distance to base IMU ‖p‖ mm	0.375	0.341	0.329

Not in this comparison

MVIS · only this tool · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Kalibr · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model
Basalt · missing · 1 row: no Kalibr rolling-shutter run; Basalt has no rolling-shutter model

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

The spreads of MVIS and Kalibr differ by 1.2×, but a gap this size appears often by chance with 10 recordings (p = 0.28). Treated as a tie. The spreads of MVIS and Basalt differ by 1.4×, but a gap this size appears often by chance with 10 recordings (p = 0.14). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.4×, but a gap this size appears often by chance with 10 recordings (p = 0.28). Treated as a tie.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 6 param.	1.15	0.94 – 1.64	0.28 (raw 0.17)	tie: gap within chance
MVIS ÷ Basalt · 6 param.	0.73	0.5 – 1	0.14 (raw 0.047)	tie: gap within chance
Kalibr ÷ Basalt · 6 param.	0.73	0.41 – 1.17	0.28 (raw 0.14)	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper MEMS imu0 · gyro Dw[0,0] scale x	0.000875	0.00131	0.00144
Looper MEMS imu0 · gyro Dw[1,1] scale y	0.000964	0.000679	0.00131
Looper MEMS imu0 · gyro Dw[2,2] scale z	0.000329	0.000342	0.000454
Looper MEMS imu0 · gyro Dw[1,0] cross	0.00122	0.00103	0.000703
Looper MEMS imu0 · gyro Dw[2,0] cross	0.000459	0.000337	0.00048
Looper MEMS imu0 · gyro Dw[2,1] cross	0.000619	0.000549	0.00111

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T11 · 10 matched recordings (MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10) · 3-cam + IMU, cam+IMU

The gap between MVIS and Kalibr is clear but under 20 % (1.2×), so it is treated as a tie. The spreads of MVIS and Basalt differ by 1.2×, but a gap this size appears often by chance with 10 recordings (p = 0.7). Treated as a tie. The spreads of Kalibr and Basalt differ by 1.01×, but a gap this size appears often by chance with 10 recordings (p = 0.92). Treated as a tie.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 3 param.	0.82	0.62 – 0.95	0.029 (raw 0.0098)	tie: gap under 20 %
MVIS ÷ Basalt · 3 param.	0.81	0.6 – 1.4	0.7 (raw 0.35)	tie: gap within chance
Kalibr ÷ Basalt · 3 param.	1.01	0.78 – 1.93	0.92	tie: gap within chance

Parameter	MVIS sd	Kalibr sd	Basalt sd
Looper MEMS imu0 · accel Da[0,0] scale x	0.000395	0.000286	0.000485
Looper MEMS imu0 · accel Da[1,1] scale y	0.000281	0.000502	0.000495
Looper MEMS imu0 · accel Da[2,2] scale z	0.00159	0.00195	0.00179

c MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Parameter tables, tool differences and run configs (Looper Insight-9)

No ground truth: ◆ marks disagreement, not who is right.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

Key and MVIS column

Swipe a table sideways for the other tools and differences.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

T4 · Camera stereo

MVIS 10 of 10 · Kalibr 10 of 10 · Basalt 10 of 10

Camera · stereo01–stereo10 · Kalibr pinhole-equi

cam0 – Looper GS cam0 · 8 rows · ◆ 3

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	335.60 ± 0.38 (10)	335.00 ± 0.19 (10)	335.33 ± 0.38 (10)	◆ +0.60	+0.33	−0.27
fy px	335.60 ± 0.37 (10)	334.97 ± 0.19 (10)	335.31 ± 0.38 (10)	◆ +0.63	+0.35	−0.28
cx px	271.99 ± 0.11 (10)	271.93 ± 0.17 (10)	272.01 ± 0.11 (10)	+0.06	+0.09	+0.03
cy px	318.49 ± 0.05 (10)	318.48 ± 0.05 (10)	318.46 ± 0.03 (10)	+0.02	−0.01	−0.03
dist k1	−0.02356 ± 0.00109 (10)	−0.02167 ± 0.00139 (10)	−0.02323 ± 0.00079 (10)	◆ −0.00189	◆ −0.00156	+0.00033
dist k2	0.01131 ± 0.00284 (10)	0.01157 ± 0.00481 (10)	0.01362 ± 0.00156 (10)	−0.00026	+0.00205	+0.00230
dist k3	−0.01022 ± 0.00381 (10)	−0.01366 ± 0.00588 (10)	−0.01356 ± 0.00177 (10)	+0.00344	+0.00010	−0.00334
dist k4	0.00123 ± 0.00170 (10)	0.00324 ± 0.00244 (10)	0.00236 ± 0.00064 (10)	−0.00200	−0.00088	+0.00113

cam1 – Looper GS cam1 · 9 rows · ◆ 6

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	335.62 ± 0.40 (10)	335.09 ± 0.20 (10)	335.41 ± 0.38 (10)	◆ +0.53	+0.32	−0.22
fy px	335.61 ± 0.40 (10)	335.07 ± 0.20 (10)	335.38 ± 0.38 (10)	◆ +0.54	+0.31	−0.23
cx px	269.29 ± 0.12 (10)	269.42 ± 0.16 (10)	269.31 ± 0.09 (10)	−0.13	−0.10	+0.02
cy px	318.27 ± 0.07 (10)	318.25 ± 0.05 (10)	318.22 ± 0.05 (10)	+0.01	−0.03	−0.04
dist k1	−0.02366 ± 0.00110 (10)	−0.02123 ± 0.00095 (10)	−0.02336 ± 0.00103 (10)	◆ −0.00243	◆ −0.00212	+0.00030
dist k2	0.01135 ± 0.00224 (10)	0.00922 ± 0.00393 (10)	0.01321 ± 0.00226 (10)	+0.00213	◆ +0.00399	+0.00186
dist k3	−0.00975 ± 0.00259 (10)	−0.00998 ± 0.00515 (10)	−0.01417 ± 0.00288 (10)	+0.00022	−0.00420	◆ −0.00442
dist k4	0.00053 ± 0.00103 (10)	0.00150 ± 0.00219 (10)	0.00304 ± 0.00126 (10)	−0.00097	+0.00154	◆ +0.00251
baseline to Looper GS cam0 mm	99.129 ± 0.176 (10)	99.188 ± 0.238 (10)	99.126 ± 0.183 (10)	−0.059	−0.062	−0.004

T11 · 3-cam + IMU

MVIS 10 of 10 · Kalibr 10 of 10, no values on Looper RS cam2 · Basalt 10 of 10, no values on Looper RS cam2

Camera + IMU · bags 1–10 · Kalibr from its own camera calibration

Read first: Kalibr's intrinsics and baselines are inputs from its own camera-only calibration; its td and lever are solved given them.

Notes

Like for like: Every tool uses the same IMU noise, measured at rest on this unit. Kalibr's td and lever are solved given its inputs. RS cam2 is MVIS only.

Note: Kalibr's camera + IMU solver is global-shutter only, so the RS cam2 is not in its runs. Every tool uses the same IMU noise, measured at rest on this unit.

Basalt: cam0 + cam1 (no rolling-shutter model, so no cam2).

cam0 – Looper GS cam0 · 11 rows · ◆ 3

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	333.93 ± 0.15 (10)	fixed 335.00 ± 0.19 (10)	334.24 ± 0.44 (10)			+0.31
fy px	333.93 ± 0.16 (10)	fixed 334.97 ± 0.19 (10)	334.19 ± 0.43 (10)			+0.26
cx px	272.18 ± 0.14 (10)	fixed 271.93 ± 0.17 (10)	271.96 ± 0.15 (10)			◆ −0.23
cy px	318.58 ± 0.07 (10)	fixed 318.48 ± 0.05 (10)	318.51 ± 0.04 (10)			−0.07
dist k1	−0.02193 ± 0.00122 (10)	fixed −0.02167 ± 0.00139 (10)	−0.02156 ± 0.00120 (10)			+0.00037
dist k2	0.01207 ± 0.00406 (10)	fixed 0.01157 ± 0.00481 (10)	0.01092 ± 0.00243 (10)			−0.00115
dist k3	−0.01359 ± 0.00479 (10)	fixed −0.01366 ± 0.00588 (10)	−0.01093 ± 0.00284 (10)			+0.00266
dist k4	0.00288 ± 0.00186 (10)	fixed 0.00324 ± 0.00244 (10)	0.00135 ± 0.00111 (10)			−0.00153
time offset td to base IMU^‡ ms	1.326 ± 0.087 (10)	−0.272 ± 0.131 (10)	0.580 ± 0.156 (10)	◆ +1.598	◆ +0.852	◆ −0.746
distance to base IMU ‖p‖ mm	43.73 ± 0.37 (10)	44.56 ± 0.51 (10)	45.05 ± 0.56 (10)	◆ −0.83	+0.49	◆ +1.33
reproj median px	0.153 ± 0.006 (10)	0.195 ± 0.008 (10)	not reported	fit

cam1 – Looper GS cam1 · 12 rows · ◆ 8

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	333.99 ± 0.12 (10)	fixed 335.09 ± 0.20 (10)	334.26 ± 0.45 (10)			+0.27
fy px	333.91 ± 0.13 (10)	fixed 335.07 ± 0.20 (10)	334.22 ± 0.44 (10)			+0.32
cx px	269.53 ± 0.12 (10)	fixed 269.42 ± 0.16 (10)	269.25 ± 0.09 (10)			◆ −0.29
cy px	318.46 ± 0.09 (10)	fixed 318.25 ± 0.05 (10)	318.31 ± 0.06 (10)			◆ −0.15
dist k1	−0.02296 ± 0.00090 (10)	fixed −0.02123 ± 0.00095 (10)	−0.02152 ± 0.00137 (10)			◆ +0.00144
dist k2	0.01788 ± 0.00329 (10)	fixed 0.00922 ± 0.00393 (10)	0.01120 ± 0.00265 (10)			◆ −0.00668
dist k3	−0.01837 ± 0.00428 (10)	fixed −0.00998 ± 0.00515 (10)	−0.01195 ± 0.00328 (10)			◆ +0.00642
dist k4	0.00396 ± 0.00176 (10)	fixed 0.00150 ± 0.00219 (10)	0.00210 ± 0.00133 (10)			◆ −0.00186
baseline to Looper GS cam0 mm	99.003 ± 0.211 (10)	fixed 99.188 ± 0.238 (10)	98.980 ± 0.191 (10)			−0.023
time offset td to base IMU^‡ ms	1.360 ± 0.092 (10)	−0.256 ± 0.135 (10)	0.580 ± 0.156 (10)	◆ +1.616	◆ +0.836	◆ −0.780
distance to base IMU ‖p‖ mm	63.96 ± 0.38 (10)	63.33 ± 0.34 (10)	63.21 ± 0.33 (10)	◆ +0.63	−0.12	◆ −0.74
reproj median px	0.161 ± 0.006 (10)	0.209 ± 0.009 (10)	not reported	fit

cam2 – Looper RS cam2 · 13 rows

Kalibr: not in its runs, which are global-shutter only.

Basalt: no rolling-shutter model.

Parameter	MVISKalibr IMU model	Kalibr	Basalt
fx px	775.21 ± 0.53 (10)	no run	no run
fy px	775.31 ± 0.49 (10)	no run	no run
cx px	960.13 ± 0.22 (10)	no run	no run
cy px	535.71 ± 0.27 (10)	no run	no run
dist k1	−0.03264 ± 0.00129 (10)	no run	no run
dist k2	0.00067 ± 0.00441 (10)	no run	no run
dist k3	−0.00223 ± 0.00569 (10)	no run	no run
dist k4	0.00072 ± 0.00238 (10)	no run	no run
readout tr ms	32.148 ± 0.231 (10)	no run	no run
baseline to Looper GS cam0 mm	49.527 ± 0.116 (10)	no run	no run
time offset td to base IMU^‡ ms	−5.673 ± 0.150 (10)	no run	no run
distance to base IMU ‖p‖ mm	22.94 ± 0.43 (10)	no run	no run
reproj median px	0.434 ± 0.016 (10)	no run	no run

base IMU – Looper MEMS imu0 · 19 rows · ◆ 4

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
gyro Dw[0,0] scale x	1.00082 ± 0.00087 (10)	0.99846 ± 0.00131 (10)	0.99882 ± 0.00144 (10)	◆ +0.00236	+0.00036	◆ −0.00200
gyro Dw[1,1] scale y	1.00000 ± 0.00096 (10)	0.99998 ± 0.00068 (10)	0.99995 ± 0.00131 (10)	+0.00002	−0.00003	−0.00005
gyro Dw[2,2] scale z	0.99572 ± 0.00033 (10)	0.99575 ± 0.00034 (10)	0.99571 ± 0.00045 (10)	−0.00003	−0.00004	−0.00001
gyro Dw[1,0] cross	0.00176 ± 0.00122 (10)	0.00175 ± 0.00103 (10)	0.00206 ± 0.00070 (10)	+0.00001	+0.00031	+0.00030
gyro Dw[2,0] cross	−0.00078 ± 0.00046 (10)	0.00127 ± 0.00034 (10)	0.00008 ± 0.00048 (10)	◆ −0.00205	◆ −0.00119	◆ +0.00086
gyro Dw[2,1] cross	−0.00166 ± 0.00062 (10)	−0.00180 ± 0.00055 (10)	−0.00176 ± 0.00111 (10)	+0.00014	+0.00004	−0.00010
accel Da[0,0] scale x	0.99618 ± 0.00040 (10)	0.99574 ± 0.00029 (10)	0.99563 ± 0.00049 (10)	◆ +0.00043	−0.00011	◆ −0.00055
accel Da[1,1] scale y	0.99681 ± 0.00028 (10)	0.99636 ± 0.00050 (10)	0.99625 ± 0.00049 (10)	+0.00045	−0.00011	◆ −0.00055
accel Da[2,2] scale z	1.00457 ± 0.00159 (10)	1.00541 ± 0.00195 (10)	1.00477 ± 0.00179 (10)	−0.00084	−0.00064	+0.00020
Tg[0,0] g-sensitivity mrad·s/m	−0.077 ± 0.075 (10)	−0.234 ± 0.108 (10)	not in model	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.029 ± 0.039 (10)	−0.246 ± 0.118 (10)	not in model	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.073 ± 0.130 (10)	0.063 ± 0.184 (10)	not in model	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.131 ± 0.099 (10)	0.804 ± 0.164 (10)	not in model	≠ model
Tg[1,1] g-sensitivity mrad·s/m	0.006 ± 0.078 (10)	0.021 ± 0.055 (10)	not in model	≠ model
Tg[1,2] g-sensitivity mrad·s/m	0.051 ± 0.162 (10)	−2.207 ± 0.370 (10)	not in model	≠ model
Tg[2,0] g-sensitivity mrad·s/m	0.139 ± 0.189 (10)	0.000 ± 0.258 (10)	not in model	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.048 ± 0.041 (10)	1.381 ± 0.082 (10)	not in model	≠ model
Tg[2,2] g-sensitivity mrad·s/m	0.001 ± 0.068 (10)	−0.018 ± 0.190 (10)	not in model	≠ model
t_ga gyro−accel ms	−5.297 ± 0.107 (10)	not in model	not in model

MVIS default IMU model (calyx) · 36 of 55 rows, model-independent only · ◆ 4

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – Looper GS cam0
fx px	333.93 ± 0.15 (10)	333.92 ± 0.16 (10)	fixed 335.00 ± 0.19 (10)		−0.01
fy px	333.93 ± 0.16 (10)	333.91 ± 0.16 (10)	fixed 334.97 ± 0.19 (10)		−0.01
cx px	272.18 ± 0.14 (10)	272.18 ± 0.13 (10)	fixed 271.93 ± 0.17 (10)		0.00
cy px	318.58 ± 0.07 (10)	318.58 ± 0.07 (10)	fixed 318.48 ± 0.05 (10)		0.00
dist k1	−0.02193 ± 0.00122 (10)	−0.02197 ± 0.00123 (10)	fixed −0.02167 ± 0.00139 (10)		−0.00004
dist k2	0.01207 ± 0.00406 (10)	0.01205 ± 0.00408 (10)	fixed 0.01157 ± 0.00481 (10)		−0.00002
dist k3	−0.01359 ± 0.00479 (10)	−0.01349 ± 0.00483 (10)	fixed −0.01366 ± 0.00588 (10)		+0.00009
dist k4	0.00288 ± 0.00186 (10)	0.00283 ± 0.00187 (10)	fixed 0.00324 ± 0.00244 (10)		−0.00005
time offset td to base IMU^‡ ms	1.326 ± 0.087 (10)	1.315 ± 0.077 (10)	−0.272 ± 0.131 (10)	◆ +1.587	−0.011
distance to base IMU ‖p‖ mm	43.73 ± 0.37 (10)	43.70 ± 0.37 (10)	44.56 ± 0.51 (10)	◆ −0.86	−0.03
reproj median px	0.153 ± 0.006 (10)	0.154 ± 0.006 (10)	0.195 ± 0.008 (10)	fit	fit
cam1 – Looper GS cam1
fx px	333.99 ± 0.12 (10)	333.99 ± 0.12 (10)	fixed 335.09 ± 0.20 (10)		0.00
fy px	333.91 ± 0.13 (10)	333.91 ± 0.13 (10)	fixed 335.07 ± 0.20 (10)		0.00
cx px	269.53 ± 0.12 (10)	269.53 ± 0.12 (10)	fixed 269.42 ± 0.16 (10)		0.00
cy px	318.46 ± 0.09 (10)	318.46 ± 0.08 (10)	fixed 318.25 ± 0.05 (10)		0.00
dist k1	−0.02296 ± 0.00090 (10)	−0.02297 ± 0.00089 (10)	fixed −0.02123 ± 0.00095 (10)		0.00000
dist k2	0.01788 ± 0.00329 (10)	0.01779 ± 0.00328 (10)	fixed 0.00922 ± 0.00393 (10)		−0.00009
dist k3	−0.01837 ± 0.00428 (10)	−0.01823 ± 0.00426 (10)	fixed −0.00998 ± 0.00515 (10)		+0.00014
dist k4	0.00396 ± 0.00176 (10)	0.00393 ± 0.00174 (10)	fixed 0.00150 ± 0.00219 (10)		−0.00003
baseline to Looper GS cam0 mm	99.003 ± 0.211 (10)	99.001 ± 0.211 (10)	fixed 99.188 ± 0.238 (10)		−0.002
time offset td to base IMU^‡ ms	1.360 ± 0.092 (10)	1.349 ± 0.080 (10)	−0.256 ± 0.135 (10)	◆ +1.606	−0.010
distance to base IMU ‖p‖ mm	63.96 ± 0.38 (10)	63.96 ± 0.37 (10)	63.33 ± 0.34 (10)	◆ +0.63	0.00
reproj median px	0.161 ± 0.006 (10)	0.162 ± 0.006 (10)	0.209 ± 0.009 (10)	fit	fit
cam2 – Looper RS cam2
fx px	775.21 ± 0.53 (10)	775.21 ± 0.51 (10)	no run		0.00
fy px	775.31 ± 0.49 (10)	775.33 ± 0.48 (10)	no run		+0.02
cx px	960.13 ± 0.22 (10)	960.12 ± 0.22 (10)	no run		−0.01
cy px	535.71 ± 0.27 (10)	535.69 ± 0.27 (10)	no run		−0.02
dist k1	−0.03264 ± 0.00129 (10)	−0.03266 ± 0.00131 (10)	no run		−0.00002
dist k2	0.00067 ± 0.00441 (10)	0.00066 ± 0.00441 (10)	no run		−0.00001
dist k3	−0.00223 ± 0.00569 (10)	−0.00217 ± 0.00568 (10)	no run		+0.00006
dist k4	0.00072 ± 0.00238 (10)	0.00069 ± 0.00237 (10)	no run		−0.00003
readout tr ms	32.148 ± 0.231 (10)	32.136 ± 0.224 (10)	no run		−0.012
baseline to Looper GS cam0 mm	49.527 ± 0.116 (10)	49.527 ± 0.117 (10)	no run		0.000
time offset td to base IMU^‡ ms	−5.673 ± 0.150 (10)	−5.687 ± 0.141 (10)	no run		−0.014
distance to base IMU ‖p‖ mm	22.94 ± 0.43 (10)	22.90 ± 0.38 (10)	no run		−0.04
reproj median px	0.434 ± 0.016 (10)	0.434 ± 0.016 (10)	no run		fit

Monado MGC

Reverb G2 headset calibration from the Monado SLAM Datasets. Example.

Monado MGC: which tool repeats more tightly, per family and group of parameters
Family	Focal + pp px	Distortion	Distance mm	Time offset ms	Cam–IMU distance mm	Gyro	Accel
T5 · camera n = 4 matched cam0 + cam1 stereo	n = 4 · too few	n = 4 · too few ^a	n = 4 · too few	—	—	—	—
T12 · cam+IMU n = 4 matched cam0 + cam1 + IMU	n = 4 · too few	n = 4 · too few ^a	n = 4 · too few	n = 4 · too few ^bc	n = 4 · too few ^b	n = 4 · too few ^bd	n = 4 · too few ^bd

T5 · camera · cam0 + cam1 stereo Too few recordings (n = 4) n = 4 matched · MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Focal + ppToo few recordings (n = 4)
DistortionToo few recordings (n = 4)^a
DistanceToo few recordings (n = 4)

T12 · cam+IMU · cam0 + cam1 + IMU Too few recordings (n = 4) n = 4 matched · MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Focal + ppToo few recordings (n = 4)
DistortionToo few recordings (n = 4)^a
DistanceToo few recordings (n = 4)
Time offsetToo few recordings (n = 4)^bc
Cam–IMU distanceToo few recordings (n = 4)^b
GyroToo few recordings (n = 4)^bd
AccelToo few recordings (n = 4)^bd

k = the looser tool's typical scatter divided by the tighter tool's (median over the group's parameters, on the recordings every compared tool ran). Click a cell for its numbers.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
b Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
c MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.
d MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Each cell is a separate comparison on the recordings every compared tool ran. Cells are not added up and nothing is ordered across datasets.

Not part of this comparison

T12 reprojection error (2 rows): fit quality, not repeatability and not accuracy
T12 IMU g-sensitivity Tg (9 rows): Kalibr absorbs it in its IMU model
T12 gyro–accel delay t_ga (1 row): MVIS only

Focal length + principal point · T5 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 stereo, camera

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 8 param.	0.69	not bounded at n = 4	1 (raw 0.38)	too few recordings
MVIS ÷ Basalt · 8 param.	0.72	not bounded at n = 4	1 (raw 0.38)	too few recordings
Kalibr ÷ Basalt · 8 param.	1.02	not bounded at n = 4	1	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 cam0 · fx px	0.0452	0.803	0.175
Reverb G2 cam0 · fy px	0.0771	0.758	0.154
Reverb G2 cam0 · cx px	0.225	0.329	0.334
Reverb G2 cam0 · cy px	0.187	0.186	0.221
Reverb G2 cam1 · fx px	0.176	0.096	0.167
Reverb G2 cam1 · fy px	0.152	0.112	0.163
Reverb G2 cam1 · cx px	0.235	0.341	0.306
Reverb G2 cam1 · cy px	0.159	0.288	0.275

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T5 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 stereo, camera

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 8 param.	0.23	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 8 param.	2.55	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 8 param.	8.92	not bounded at n = 4	0.38 (raw 0.12)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 cam0 · dist k1	0.00215	0.0313	0.00202
Reverb G2 cam0 · dist k2	0.00772	0.111	0.00381
Reverb G2 cam0 · dist k3	0.0113	0.15	0.00445
Reverb G2 cam0 · dist k4	0.00538	0.0678	0.0021
Reverb G2 cam1 · dist k1	0.0031	0.00473	0.00125
Reverb G2 cam1 · dist k2	0.0136	0.0166	0.00323
Reverb G2 cam1 · dist k3	0.0201	0.022	0.00457
Reverb G2 cam1 · dist k4	0.00944	0.00978	0.00224

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T5 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 stereo, camera

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its camera-only run.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 1 param.	1.19	not bounded at n = 4	1 (raw 0.62)	too few recordings
MVIS ÷ Basalt · 1 param.	1.28	not bounded at n = 4	1 (raw 0.5)	too few recordings
Kalibr ÷ Basalt · 1 param.	1.07	not bounded at n = 4	0.75 (raw 0.25)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 cam1 · baseline to Reverb G2 cam0 mm	0.0513	0.0429	0.0402

Repeatability only — a tool can repeat well and still be biased.

Focal length + principal point · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	0.92	not bounded at n = 4	0.62	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
Reverb G2 cam0 · fx px	0.144	0.143	1.01
Reverb G2 cam0 · fy px	0.145	0.147	0.99
Reverb G2 cam0 · cx px	0.0494	0.122	0.40
Reverb G2 cam0 · cy px	0.056	0.0653	0.86
Reverb G2 cam1 · fx px	0.161	0.138	1.17
Reverb G2 cam1 · fy px	0.136	0.132	1.03
Reverb G2 cam1 · cx px	0.0952	0.139	0.68
Reverb G2 cam1 · cy px	0.0899	0.107	0.84

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Distortion coefficients · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 8 param.	1.05	not bounded at n = 4	0.75	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
Reverb G2 cam0 · dist k1	0.00149	0.00183	0.81
Reverb G2 cam0 · dist k2	0.00279	0.00581	0.48
Reverb G2 cam0 · dist k3	0.00294	0.00819	0.36
Reverb G2 cam0 · dist k4	0.00125	0.00364	0.34
Reverb G2 cam1 · dist k1	0.00203	0.0015	1.35
Reverb G2 cam1 · dist k2	0.00744	0.00467	1.59
Reverb G2 cam1 · dist k3	0.00992	0.00597	1.66
Reverb G2 cam1 · dist k4	0.00441	0.00246	1.79

Not in this comparison

Kalibr · input, not compared · 8 rows: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

a Coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.

Repeatability only — a tool can repeat well and still be biased.

Camera–camera distance · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr input · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision	Leave one out
MVIS ÷ Basalt · 1 param.	3.11	not bounded at n = 4	0.38	too few recordings

Parameter	MVIS sd	Basalt sd	Ratio
Reverb G2 cam1 · baseline to Reverb G2 cam0 mm	0.0382	0.0123	3.11

Not in this comparison

Kalibr · input, not compared · 1 row: Kalibr's camera+IMU run starts from these values, exported from MVIS for each recording, and keeps them fixed. Their spread copies the input, so it is not Kalibr's repeatability.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU time offset · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.49	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 2 param.	1.61	not bounded at n = 4	0.38 (raw 0.25)	too few recordings
Kalibr ÷ Basalt · 2 param.	3.27	not bounded at n = 4	0.38 (raw 0.12)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 cam0 · time offset td to base IMU ms	0.118	0.246	0.0738
Reverb G2 cam1 · time offset td to base IMU ms	0.118	0.238	0.0738

b Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
c MVIS also fits a gyro–accel delay (t_ga) that Kalibr and Basalt do not, so its td is not quite the same quantity.

Repeatability only — a tool can repeat well and still be biased.

Camera–IMU distance · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its default IMU model.

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 2 param.	0.12	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 2 param.	0.19	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 2 param.	1.56	not bounded at n = 4	0.62	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 cam0 · distance to base IMU ‖p‖ mm	0.0358	0.11	0.173
Reverb G2 cam1 · distance to base IMU ‖p‖ mm	0.042	0.968	0.251

b Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.

Repeatability only — a tool can repeat well and still be biased.

Gyro scale + cross-axis · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 6 param.	0.01	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 6 param.	1.07	not bounded at n = 4	0.62	too few recordings
Kalibr ÷ Basalt · 6 param.	93.45	not bounded at n = 4	0.38 (raw 0.12)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 IMU · gyro Dw[0,0] scale x	0.000705	0.0585	0.0006
Reverb G2 IMU · gyro Dw[1,1] scale y	0.00103	0.0607	0.000503
Reverb G2 IMU · gyro Dw[2,2] scale z	0.000169	0.0162	0.00023
Reverb G2 IMU · gyro Dw[1,0] cross	0.000373	0.0345	0.000385
Reverb G2 IMU · gyro Dw[2,0] cross	0.000109	0.0264	0.000415
Reverb G2 IMU · gyro Dw[2,1] cross	0.000662	0.0489	0.00046

b Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
d MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Accel scale · T12 · 4 matched recordings (MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4) · cam0 + cam1 + IMU, cam+IMU

Only 4 recordings hold values from every compared tool. With 4 recordings the smallest possible p is 0.12, so no gap can be decided. Ratios are shown for reference only.

MVIS here: its run with Kalibr's IMU model (K-IMU).

Pair	R	95 % interval	p (Holm)	Decision
MVIS ÷ Kalibr · 3 param.	0.02	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
MVIS ÷ Basalt · 3 param.	0.13	not bounded at n = 4	0.38 (raw 0.12)	too few recordings
Kalibr ÷ Basalt · 3 param.	21.48	not bounded at n = 4	0.38 (raw 0.12)	too few recordings

Parameter	MVIS sd	Kalibr sd	Basalt sd
Reverb G2 IMU · accel Da[0,0] scale x	0.000102	0.0175	0.000814
Reverb G2 IMU · accel Da[1,1] scale y	0.000394	0.0126	0.00393
Reverb G2 IMU · accel Da[2,2] scale z	0.000882	0.0404	0.000197

b Kalibr starts from MVIS camera values for each recording; its spread here includes the spread of that input.
d MVIS here uses Kalibr's IMU model, so the gyro and accel terms mean the same thing in both tools.

Repeatability only — a tool can repeat well and still be biased.

Parameter tables, tool differences and run configs (Monado MGC)

No ground truth: ◆ marks disagreement, not who is right.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

Key and MVIS column

Swipe a table sideways for the other tools and differences.

On camera + IMU families the MVIS column uses Kalibr's IMU model, not the MVIS default. The default fills fewer rows and is in a closed fold under each family.

T5 · cam0 + cam1 stereo

MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Camera · stereo01 rounds 1–4 · Kalibr pinhole-equi, cam0 + cam1 jointly

cam0 – Reverb G2 cam0 · 8 rows · ◆ 1

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	270.01 ± 0.05 (4)	269.92 ± 0.80 (4)	269.87 ± 0.17 (4)	+0.09	−0.05	−0.15
fy px	270.15 ± 0.08 (4)	270.02 ± 0.76 (4)	269.92 ± 0.15 (4)	+0.13	−0.10	◆ −0.23
cx px	322.44 ± 0.22 (4)	322.46 ± 0.33 (4)	322.41 ± 0.33 (4)	−0.03	−0.05	−0.02
cy px	229.09 ± 0.19 (4)	229.12 ± 0.19 (4)	229.10 ± 0.22 (4)	−0.03	−0.02	+0.01
dist k1	0.08259 ± 0.00215 (4)	0.07701 ± 0.03126 (4)	0.08125 ± 0.00202 (4)	+0.00558	+0.00424	−0.00133
dist k2	−0.04528 ± 0.00772 (4)	−0.02100 ± 0.11101 (4)	−0.03825 ± 0.00381 (4)	−0.02428	−0.01726	+0.00702
dist k3	0.23454 ± 0.01129 (4)	0.20501 ± 0.14995 (4)	0.22596 ± 0.00445 (4)	+0.02952	+0.02095	−0.00857
dist k4	−0.13254 ± 0.00538 (4)	−0.12148 ± 0.06777 (4)	−0.12902 ± 0.00210 (4)	−0.01106	−0.00753	+0.00353

cam1 – Reverb G2 cam1 · 9 rows · ◆ 4

Parameter	MVIS	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	268.76 ± 0.18 (4)	268.72 ± 0.10 (4)	268.65 ± 0.17 (4)	+0.04	−0.07	−0.11
fy px	268.79 ± 0.15 (4)	268.74 ± 0.11 (4)	268.66 ± 0.16 (4)	+0.05	−0.08	−0.13
cx px	323.71 ± 0.23 (4)	323.55 ± 0.34 (4)	323.50 ± 0.31 (4)	+0.16	−0.05	−0.21
cy px	236.66 ± 0.16 (4)	237.05 ± 0.29 (4)	237.04 ± 0.27 (4)	◆ −0.39	−0.01	◆ +0.38
dist k1	0.08219 ± 0.00310 (4)	0.07818 ± 0.00473 (4)	0.07953 ± 0.00125 (4)	+0.00401	+0.00135	−0.00266
dist k2	−0.05120 ± 0.01356 (4)	−0.03220 ± 0.01665 (4)	−0.04379 ± 0.00323 (4)	◆ −0.01900	−0.01159	+0.00741
dist k3	0.24337 ± 0.02005 (4)	0.21896 ± 0.02197 (4)	0.23718 ± 0.00457 (4)	◆ +0.02441	+0.01822	−0.00619
dist k4	−0.13566 ± 0.00944 (4)	−0.12585 ± 0.00978 (4)	−0.13368 ± 0.00224 (4)	◆ −0.00982	−0.00783	+0.00198
baseline to Reverb G2 cam0 mm	108.434 ± 0.051 (4)	108.424 ± 0.043 (4)	108.416 ± 0.040 (4)	+0.010	−0.008	−0.018

T12 · cam0 + cam1 + IMU

MVIS 4 of 4 · Kalibr 4 of 4 · Basalt 4 of 4

Camera + IMU · imucalib01 rounds 1–4 · Kalibr scale-misalignment vs MVIS (kalibr model)

Read first: Kalibr's intrinsics and baselines are inputs from an MVIS export; its td and lever are solved given them.

Notes

Note: the camera + IMU twin of T5 (same pair, same 4 sessions). Both tools start from the camera calibration of the same round: MVIS from T5 camera-only (refined in cam-IMU), Kalibr from the MVIS T12 export (fixed).

Basalt: the kb4 rows, seeded from Basalt's T5 round.

cam0 – Reverb G2 cam0 · 11 rows · ◆ 8

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	269.83 ± 0.14 (4)	fixed 269.88 ± 0.12 (4)	269.71 ± 0.14 (4)			−0.13
fy px	270.09 ± 0.14 (4)	fixed 270.08 ± 0.15 (4)	269.89 ± 0.15 (4)			◆ −0.20
cx px	322.55 ± 0.05 (4)	fixed 322.62 ± 0.06 (4)	322.44 ± 0.12 (4)			−0.11
cy px	229.25 ± 0.06 (4)	fixed 229.22 ± 0.06 (4)	229.16 ± 0.07 (4)			◆ −0.09
dist k1	0.08374 ± 0.00149 (4)	fixed 0.08285 ± 0.00123 (4)	0.08021 ± 0.00183 (4)			◆ −0.00354
dist k2	−0.04770 ± 0.00279 (4)	fixed −0.04607 ± 0.00447 (4)	−0.03643 ± 0.00581 (4)			◆ +0.01128
dist k3	0.23499 ± 0.00294 (4)	fixed 0.23258 ± 0.00600 (4)	0.22293 ± 0.00819 (4)			◆ −0.01206
dist k4	−0.13155 ± 0.00124 (4)	fixed −0.13055 ± 0.00271 (4)	−0.12743 ± 0.00364 (4)			◆ +0.00412
time offset td to base IMU^‡ ms	−1.261 ± 0.115 (4)	−0.203 ± 0.246 (4)	−0.436 ± 0.074 (4)	◆ −1.058	−0.232	◆ +0.825
distance to base IMU ‖p‖ mm	53.89 ± 0.04 (4)	53.73 ± 0.11 (4)	53.67 ± 0.17 (4)	◆ +0.16	−0.06	◆ −0.22
reproj median px	0.117 ± 0.005 (4)	0.142 ± 0.064 (4)	not reported	fit

cam1 – Reverb G2 cam1 · 12 rows · ◆ 3

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
fx px	268.63 ± 0.16 (4)	fixed 268.71 ± 0.21 (4)	268.50 ± 0.14 (4)			−0.13
fy px	268.62 ± 0.14 (4)	fixed 268.71 ± 0.18 (4)	268.49 ± 0.13 (4)			−0.13
cx px	323.62 ± 0.10 (4)	fixed 323.67 ± 0.12 (4)	323.55 ± 0.14 (4)			−0.07
cy px	236.95 ± 0.09 (4)	fixed 236.98 ± 0.10 (4)	237.10 ± 0.11 (4)			◆ +0.16
dist k1	0.07967 ± 0.00203 (4)	fixed 0.07948 ± 0.00166 (4)	0.07921 ± 0.00150 (4)			−0.00046
dist k2	−0.04150 ± 0.00744 (4)	fixed −0.04137 ± 0.00538 (4)	−0.03966 ± 0.00467 (4)			+0.00183
dist k3	0.22776 ± 0.00992 (4)	fixed 0.22814 ± 0.00677 (4)	0.22831 ± 0.00597 (4)			+0.00055
dist k4	−0.12830 ± 0.00441 (4)	fixed −0.12854 ± 0.00305 (4)	−0.12953 ± 0.00246 (4)			−0.00123
baseline to Reverb G2 cam0 mm	108.500 ± 0.038 (4)	fixed 108.507 ± 0.054 (4)	108.408 ± 0.012 (4)			◆ −0.092
time offset td to base IMU^‡ ms	−1.265 ± 0.115 (4)	−0.213 ± 0.238 (4)	−0.436 ± 0.074 (4)	◆ −1.052	−0.223	◆ +0.829
distance to base IMU ‖p‖ mm	58.46 ± 0.04 (4)	58.31 ± 0.97 (4)	58.31 ± 0.25 (4)	+0.15	−0.01	−0.16
reproj median px	0.121 ± 0.009 (4)	0.140 ± 0.066 (4)	not reported	fit

base IMU – Reverb G2 IMU · 19 rows · ◆ 5

Parameter	MVISKalibr IMU model	Kalibr	Basalt	MVIS− Kalibr	Basalt− Kalibr	Basalt− MVIS
gyro Dw[0,0] scale x	1.00125 ± 0.00070 (4)	1.00286 ± 0.05848 (4)	1.00135 ± 0.00060 (4)	−0.00161	−0.00151	+0.00010
gyro Dw[1,1] scale y	1.00004 ± 0.00103 (4)	1.00248 ± 0.06067 (4)	1.00129 ± 0.00050 (4)	−0.00244	−0.00119	◆ +0.00125
gyro Dw[2,2] scale z	1.00030 ± 0.00017 (4)	1.00037 ± 0.01617 (4)	0.99974 ± 0.00023 (4)	−0.00007	−0.00063	◆ −0.00056
gyro Dw[1,0] cross	−0.00007 ± 0.00037 (4)	0.00012 ± 0.03447 (4)	0.00005 ± 0.00038 (4)	−0.00019	−0.00007	+0.00012
gyro Dw[2,0] cross	0.00004 ± 0.00011 (4)	0.00046 ± 0.02642 (4)	0.00025 ± 0.00042 (4)	−0.00042	−0.00021	+0.00021
gyro Dw[2,1] cross	−0.00039 ± 0.00066 (4)	−0.00066 ± 0.04892 (4)	0.00080 ± 0.00046 (4)	+0.00027	+0.00146	◆ +0.00119
accel Da[0,0] scale x	1.00779 ± 0.00010 (4)	1.00721 ± 0.01747 (4)	1.00679 ± 0.00081 (4)	+0.00058	−0.00042	◆ −0.00100
accel Da[1,1] scale y	1.01189 ± 0.00039 (4)	1.01291 ± 0.01257 (4)	1.01173 ± 0.00393 (4)	−0.00102	−0.00118	−0.00016
accel Da[2,2] scale z	1.00897 ± 0.00088 (4)	1.00999 ± 0.04037 (4)	1.00655 ± 0.00020 (4)	−0.00102	−0.00344	◆ −0.00242
Tg[0,0] g-sensitivity mrad·s/m	−0.039 ± 0.080 (4)	−0.020 ± 5.076 (4)	not in model	≠ model
Tg[0,1] g-sensitivity mrad·s/m	0.063 ± 0.072 (4)	0.510 ± 1.130 (4)	not in model	≠ model
Tg[0,2] g-sensitivity mrad·s/m	−0.049 ± 0.059 (4)	−2.304 ± 11.510 (4)	not in model	≠ model
Tg[1,0] g-sensitivity mrad·s/m	−0.007 ± 0.066 (4)	−0.576 ± 0.875 (4)	not in model	≠ model
Tg[1,1] g-sensitivity mrad·s/m	−0.004 ± 0.077 (4)	−0.284 ± 4.087 (4)	not in model	≠ model
Tg[1,2] g-sensitivity mrad·s/m	−0.007 ± 0.114 (4)	−0.452 ± 17.028 (4)	not in model	≠ model
Tg[2,0] g-sensitivity mrad·s/m	−0.030 ± 0.045 (4)	1.674 ± 3.397 (4)	not in model	≠ model
Tg[2,1] g-sensitivity mrad·s/m	−0.048 ± 0.066 (4)	−0.415 ± 0.838 (4)	not in model	≠ model
Tg[2,2] g-sensitivity mrad·s/m	0.005 ± 0.080 (4)	0.074 ± 4.868 (4)	not in model	≠ model
t_ga gyro−accel ms	2.517 ± 0.017 (4)	not in model	not in model

MVIS default IMU model (calyx) · 23 of 42 rows, model-independent only · ◆ 3

Parameter	MVISKalibr IMU model	MVIS default	Kalibr	default− Kalibr	default− Kalibr IMU model
cam0 – Reverb G2 cam0
fx px	269.83 ± 0.14 (4)	269.83 ± 0.14 (4)	fixed 269.88 ± 0.12 (4)		0.00
fy px	270.09 ± 0.14 (4)	270.09 ± 0.14 (4)	fixed 270.08 ± 0.15 (4)		0.00
cx px	322.55 ± 0.05 (4)	322.55 ± 0.05 (4)	fixed 322.62 ± 0.06 (4)		0.00
cy px	229.25 ± 0.06 (4)	229.25 ± 0.06 (4)	fixed 229.22 ± 0.06 (4)		0.00
dist k1	0.08374 ± 0.00149 (4)	0.08374 ± 0.00149 (4)	fixed 0.08285 ± 0.00123 (4)		0.00000
dist k2	−0.04770 ± 0.00279 (4)	−0.04770 ± 0.00279 (4)	fixed −0.04607 ± 0.00447 (4)		0.00000
dist k3	0.23499 ± 0.00294 (4)	0.23499 ± 0.00294 (4)	fixed 0.23258 ± 0.00600 (4)		0.00000
dist k4	−0.13155 ± 0.00124 (4)	−0.13155 ± 0.00125 (4)	fixed −0.13055 ± 0.00271 (4)		0.00000
time offset td to base IMU^‡ ms	−1.261 ± 0.115 (4)	−1.261 ± 0.118 (4)	−0.203 ± 0.246 (4)	◆ −1.058	0.000
distance to base IMU ‖p‖ mm	53.89 ± 0.04 (4)	53.89 ± 0.04 (4)	53.73 ± 0.11 (4)	◆ +0.16	0.00
reproj median px	0.117 ± 0.005 (4)	0.117 ± 0.004 (4)	0.142 ± 0.064 (4)	fit	fit
cam1 – Reverb G2 cam1
fx px	268.63 ± 0.16 (4)	268.63 ± 0.16 (4)	fixed 268.71 ± 0.21 (4)		0.00
fy px	268.62 ± 0.14 (4)	268.62 ± 0.14 (4)	fixed 268.71 ± 0.18 (4)		0.00
cx px	323.62 ± 0.10 (4)	323.62 ± 0.10 (4)	fixed 323.67 ± 0.12 (4)		0.00
cy px	236.95 ± 0.09 (4)	236.95 ± 0.09 (4)	fixed 236.98 ± 0.10 (4)		0.00
dist k1	0.07967 ± 0.00203 (4)	0.07967 ± 0.00203 (4)	fixed 0.07948 ± 0.00166 (4)		0.00000
dist k2	−0.04150 ± 0.00744 (4)	−0.04150 ± 0.00744 (4)	fixed −0.04137 ± 0.00538 (4)		0.00000
dist k3	0.22776 ± 0.00992 (4)	0.22776 ± 0.00992 (4)	fixed 0.22814 ± 0.00677 (4)		0.00000
dist k4	−0.12830 ± 0.00441 (4)	−0.12830 ± 0.00441 (4)	fixed −0.12854 ± 0.00305 (4)		0.00000
baseline to Reverb G2 cam0 mm	108.500 ± 0.038 (4)	108.500 ± 0.038 (4)	fixed 108.507 ± 0.054 (4)		0.000
time offset td to base IMU^‡ ms	−1.265 ± 0.115 (4)	−1.265 ± 0.118 (4)	−0.213 ± 0.238 (4)	◆ −1.053	0.000
distance to base IMU ‖p‖ mm	58.46 ± 0.04 (4)	58.46 ± 0.04 (4)	58.31 ± 0.97 (4)	+0.15	0.00
reproj median px	0.121 ± 0.009 (4)	0.121 ± 0.009 (4)	0.140 ± 0.066 (4)	fit	fit

About these results

How the verdicts are made

For each family and group of parameters, we decide a verdict in five steps:

We take only the recordings every compared tool ran.
For each tool, we measure the scatter (standard deviation) of each parameter across those recordings.
We compare two tools by the median ratio of their scatters.
A tool is marked more repeatable only if both of these hold:
- its scatter is at least 20 % smaller, and
- a gap that large would rarely appear (under 5 %, Holm-adjusted within the cell) if we swapped the two tools’ results recording by recording.
Otherwise it is a tie: either the gap is within chance, or it is under 20 %.

Notes on the test:

Too few recordings: with fewer than 6 shared recordings, no gap can pass this test (the smallest possible p is 2/2ⁿ). Those cells say “too few recordings”.
Not compared: values a tool received as input, and rows where the tools’ models differ.
Distortion coefficients trade off against each other, so their scatter can overstate differences in the lens mapping.
Many tests: each cell is a separate 5 % test. Across the 27 tested cells on this page, of which 14 name a tool, one or two calls may be chance. The intervals are not adjusted for multiple tests.

Agreement between tools

The parameter tables also show the differences between the tools’ medians. No tool is the reference.

Disagreement: two tools disagree where their medians differ by more than either tool’s scatter.
◆ marks disagreement, not an error in either tool.
‡ on a time offset: MVIS splits the gyro (td) and accel (td + t_ga) timelines, so Basalt − Kalibr is the like-for-like check there. Here td is the camera–IMU time offset, and t_ga is the time offset between the gyroscope and accelerometer channels.

How to read the parameter tables

How to read this

Spread shows repeatability across recordings, not accuracy. A tool can repeat well and still be biased.
448.50 ± 0.48 (10): median ± sd across the 10 recordings that tool solved. With 3 or 4 recordings the sd itself is rough.
(3/15): this tool has 3 of the family's 15 recordings.
fixed: Kalibr doesn't estimate this value. Its camera + IMU runs take intrinsics and baselines as input, from an MVIS export or, on T6 and T11, from its own camera-only calibration; T7 ELP-L readout is held at 0. The spread here isn't repeatability.
shown: the tool uses a different model here, so the value has no difference next to it.
other model: Kalibr's coefficients here belong to a different lens model.
not in model, not reported, no run, no data, no RS model, not in one run: the tool has no value for this row, and the words say why. no run means the tool never ran with this sensor; not in one run means it ran with each sensor, but no single run holds both ends of the baseline. not run, in the recordings line: the tool has no run for the family.
Differences: the column header gives the subtraction, using stored medians. No tool is the reference.
◆: the difference is larger than the larger of the two sds. It marks disagreement, not an error in either tool.
≠n: the two medians come from different numbers of recordings.
≠ model: the tools' models differ, so no difference is shown. fit: reprojection error measures fit and isn't comparable across models.
‡ td: MVIS splits the IMU into gyro (td) and accel (td + t_ga); Kalibr and Basalt have one timeline. Basalt − Kalibr is the like-for-like check.
Each family folds; the first is open. Click a family's title to open or close it.

Family numbers

Each family is one test in the MVIS regression suite. On 2026-11-02 the numbers changed so the VI-Rig datasets come first. Older material uses the old numbers.

New	Old
T1	T3
T2	T4
T3	T1
T4	T2
T5	T5
T6	T11
T7	T8
T8	T12
T9	T6
T10	T7
T11	T10
T12	T9

Methodology · Submit results · HF dataset · Space