VI-Rig: 5 IMUs + 4 cameras
Repeatability, not accuracy. There is no ground truth: a small spread shows that a tool repeats, not that it is right, and a difference shows that two tools disagree, not which one is wrong. The MVIS authors made every run; only MVIS was run on this set.
Repeatability
Section titled “Repeatability”This study has no repeatability cells: a cell compares how much each tool's estimates scatter, and only MVIS was run here. The spread of each MVIS value is in Parameters by sensor.
V5 · camera + IMU
Section titled “V5 · camera + IMU”4 t265 fisheye cameras + 5 IMUs · 7 recordings
Read first: MVIS only: Kalibr and Basalt were not run on this set, so nothing is compared. The spread shows how much MVIS repeats over the 7 recordings, not that it is right.
Coverage
Section titled “Coverage”- MVIS 7 of 7
- Kalibr 0 of 7 not run on this set
- Basalt 0 of 7 not run on this set
Fit: reprojection median, px (fit of each tool's own model, not accuracy)
| Camera | MVIS |
|---|---|
| t265 A fisheye 1 | 0.208 |
| t265 A fisheye 2 | 0.219 |
| t265 B fisheye 1 | 0.206 |
| t265 B fisheye 2 | 0.228 |
Parameters by sensor
Section titled “Parameters by sensor”9 sensors, 101 parameter charts (open to see each tool's spread)
fx px axis 287 ± 0.969
fy px axis 286 ± 0.74
cx px axis 415 ± 0.707
cy px axis 404 ± 0.626
dist k1 axis −0.00758 ± 0.00523
dist k2 axis 0.042 ± 0.0115
dist k3 axis −0.0373 ± 0.0143
dist k4 axis 0.00593 ± 0.00529
time offset td to base IMU ms axis −0.337 ± 0.176
distance to base IMU ‖p‖ mm axis 194 ± 1.2
fx px axis 286 ± 0.871
fy px axis 285 ± 0.814
cx px axis 414 ± 0.639
cy px axis 399 ± 0.688
dist k1 axis −0.0064 ± 0.00444
dist k2 axis 0.045 ± 0.00977
dist k3 axis −0.0429 ± 0.0115
dist k4 axis 0.0081 ± 0.00402
baseline to t265 A fisheye 1 mm axis 63.7 ± 0.193
time offset td to base IMU ms axis −0.332 ± 0.201
distance to base IMU ‖p‖ mm axis 148 ± 1.04
fx px axis 286 ± 0.667
fy px axis 285 ± 0.436
cx px axis 417 ± 0.293
cy px axis 400 ± 0.783
dist k1 axis −0.00711 ± 0.00533
dist k2 axis 0.0459 ± 0.0162
dist k3 axis −0.0442 ± 0.0183
dist k4 axis 0.00844 ± 0.0066
baseline to t265 A fisheye 1 mm axis 24.5 ± 1.17
time offset td to base IMU ms axis −0.379 ± 0.204
distance to base IMU ‖p‖ mm axis 201 ± 1.12
fx px axis 286 ± 0.519
fy px axis 286 ± 0.462
cx px axis 413 ± 0.405
cy px axis 399 ± 0.763
dist k1 axis −0.00818 ± 0.00608
dist k2 axis 0.0426 ± 0.0149
dist k3 axis −0.0389 ± 0.0125
dist k4 axis 0.0067 ± 0.00454
baseline to t265 A fisheye 1 mm axis 66.9 ± 0.766
time offset td to base IMU ms axis −0.366 ± 0.203
distance to base IMU ‖p‖ mm axis 158 ± 0.911
gyro Dw[0,0] scale x axis 1 ± 0.00229 · line at 1
gyro Dw[1,1] scale y axis 0.998 ± 0.0016
gyro Dw[2,2] scale z axis 0.998 ± 0.00365 · line at 1
gyro Dw[1,0] cross axis −0.000538 ± 0.00559
gyro Dw[2,0] cross axis −0.00113 ± 0.00571
gyro Dw[2,1] cross axis −0.000502 ± 0.00297
accel Da[0,0] scale x axis 1 ± 0.00215 · line at 1
accel Da[1,1] scale y axis 1 ± 0.00387 · line at 1
accel Da[2,2] scale z axis 0.996 ± 0.00278
t_ga gyro−accel ms axis −0.557 ± 0.369
gyro Dw[0,0] scale x axis 1 ± 0.00099
gyro Dw[1,1] scale y axis 1 ± 0.001 · line at 1
gyro Dw[2,2] scale z axis 1 ± 0.00117 · line at 1
gyro Dw[1,0] cross axis −0.000885 ± 0.0025
gyro Dw[2,0] cross axis −0.00252 ± 0.00433
gyro Dw[2,1] cross axis 0.00591 ± 0.00255
accel Da[0,0] scale x axis 0.997 ± 0.00231
accel Da[1,1] scale y axis 1 ± 0.00145 · line at 1
accel Da[2,2] scale z axis 1 ± 0.0029 · line at 1
time delay td to base IMU ms axis −0.555 ± 0.227
distance to base IMU ‖p‖ mm axis 131 ± 0.861
t_ga gyro−accel ms axis 0.12 ± 0.289 · line at 0
gyro Dw[0,0] scale x axis 1 ± 0.00166 · line at 1
gyro Dw[1,1] scale y axis 1 ± 0.00174 · line at 1
gyro Dw[2,2] scale z axis 0.998 ± 0.00247 · line at 1
gyro Dw[1,0] cross axis −0.00328 ± 0.00402
gyro Dw[2,0] cross axis −0.00196 ± 0.00427
gyro Dw[2,1] cross axis 0.00386 ± 0.00384
accel Da[0,0] scale x axis 0.999 ± 0.00571 · line at 1
accel Da[1,1] scale y axis 1 ± 0.00255 · line at 1
accel Da[2,2] scale z axis 0.999 ± 0.00364 · line at 1
time delay td to base IMU ms axis −2.16 ± 0.199
distance to base IMU ‖p‖ mm axis 94.7 ± 0.89
t_ga gyro−accel ms axis −0.0498 ± 0.208 · line at 0
gyro Dw[0,0] scale x axis 0.994 ± 0.000925
gyro Dw[1,1] scale y axis 0.998 ± 0.00099
gyro Dw[2,2] scale z axis 1 ± 0.000605
gyro Dw[1,0] cross axis 0.00163 ± 0.00205
gyro Dw[2,0] cross axis 0.023 ± 0.00207
gyro Dw[2,1] cross axis 0.00303 ± 0.00297
accel Da[0,0] scale x axis 1.02 ± 0.00137
accel Da[1,1] scale y axis 1.01 ± 0.00311
accel Da[2,2] scale z axis 1.02 ± 0.00165
time delay td to base IMU ms axis −7.17 ± 0.198
distance to base IMU ‖p‖ mm axis 181 ± 1.02
t_ga gyro−accel ms axis −6.59 ± 0.207
gyro Dw[0,0] scale x axis 0.997 ± 0.00106
gyro Dw[1,1] scale y axis 0.998 ± 0.000883
gyro Dw[2,2] scale z axis 0.997 ± 0.000733
gyro Dw[1,0] cross axis 0.0132 ± 0.00296
gyro Dw[2,0] cross axis 0.0243 ± 0.00261
gyro Dw[2,1] cross axis 0.00236 ± 0.00259
accel Da[0,0] scale x axis 1.01 ± 0.00121
accel Da[1,1] scale y axis 1.01 ± 0.00218
accel Da[2,2] scale z axis 1.02 ± 0.00207
time delay td to base IMU ms axis −7.26 ± 0.227
distance to base IMU ‖p‖ mm axis 189 ± 0.862
t_ga gyro−accel ms axis −6.65 ± 0.278
Whisker = lowest to highest value over the tool's solved recordings; box = median ± sd; tick = median. Min and max come from the per-recording values behind each record's median; only these two numbers are published, not the values. Rotations are not shown: the records do not hold them yet.
Model notes and how each tool was run
Section titled “Model notes and how each tool was run”4 notes on models, inputs and run settings
- t_ga, the gyro–accel delay, is estimated by MVIS only; Kalibr and Basalt have no such term, so MVIS's td is not quite the same quantity.
- Basalt not run: Basalt was not run on this set.
- Like for like: One MVIS configuration for all 7 recordings (the fast settings of the 4-IMU rows); there is no other tool to match.
- Note: Intrinsics start from the 7-recording median of the camera-only solves and are refined; fast settings: the three cameras besides cam0 enter the solve at 4 Hz.
V5c · camera only
Section titled “V5c · camera only”4 t265 fisheye cameras — camera only · 7 recordings
Read first: MVIS only: Kalibr and Basalt were not run on this set, so nothing is compared. The spread shows how much MVIS repeats over the 7 recordings, not that it is right.
Coverage
Section titled “Coverage”- MVIS 7 of 7
- Kalibr 0 of 7 not run on this set
- Basalt 0 of 7 not run on this set
Fit: reprojection median, px (fit of each tool's own model, not accuracy)
| Camera | MVIS |
|---|---|
| t265 A fisheye 1 | 0.146 |
| t265 A fisheye 2 | 0.146 |
| t265 B fisheye 1 | 0.164 |
| t265 B fisheye 2 | 0.171 |
Parameters by sensor
Section titled “Parameters by sensor”4 sensors, 34 parameter charts (open to see each tool's spread)
fx px axis 286 ± 1.13
fy px axis 285 ± 0.741
cx px axis 415 ± 0.614
cy px axis 405 ± 0.482
dist k1 axis −0.0071 ± 0.00354
dist k2 axis 0.0444 ± 0.00683
dist k3 axis −0.0425 ± 0.00917
dist k4 axis 0.00825 ± 0.00334
fx px axis 286 ± 1.02
fy px axis 285 ± 0.697
cx px axis 414 ± 0.637
cy px axis 399 ± 0.328
dist k1 axis −0.0061 ± 0.00276
dist k2 axis 0.0427 ± 0.00958
dist k3 axis −0.0397 ± 0.00718
dist k4 axis 0.00705 ± 0.00258
baseline to t265 A fisheye 1 mm axis 63.6 ± 0.192
fx px axis 286 ± 0.949
fy px axis 285 ± 0.706
cx px axis 416 ± 0.667
cy px axis 400 ± 0.651
dist k1 axis −0.0078 ± 0.0053
dist k2 axis 0.0479 ± 0.00906
dist k3 axis −0.0456 ± 0.0073
dist k4 axis 0.00881 ± 0.00239
fx px axis 286 ± 0.983
fy px axis 285 ± 0.755
cx px axis 413 ± 0.699
cy px axis 400 ± 0.572
dist k1 axis −0.00682 ± 0.00283
dist k2 axis 0.0466 ± 0.0119
dist k3 axis −0.0427 ± 0.0161
dist k4 axis 0.00813 ± 0.00978
baseline to t265 B fisheye 1 mm axis 63.1 ± 0.364
Whisker = lowest to highest value over the tool's solved recordings; box = median ± sd; tick = median. Min and max come from the per-recording values behind each record's median; only these two numbers are published, not the values. Rotations are not shown: the records do not hold them yet.
Model notes and how each tool was run
Section titled “Model notes and how each tool was run”4 notes on models, inputs and run settings
- No charted value is an input or comes from another model.
- Basalt not run: Basalt was not run on this set.
- Like for like: One MVIS configuration for all 7 recordings (the fast settings of the 4-IMU rows); there is no other tool to match.
- Note: The two T265 units are not hardware-synced to each other, so each camera-only solve holds one pair; the camera + IMU family ties all four cameras together through the IMUs.
Reading these results
Section titled “Reading these results”- Two T265 units, one board. Each pair is hardware-synced inside its unit, but the two units are not synced to each other, so each camera-only solve holds one pair (time offset off). The camera + IMU solves hold all four cameras and five IMUs, and their intrinsics start from the 7-recording median of the camera-only solves.
- T265 IMU timing. The T265 IMUs' gyro–accel delay t_ga is about −6.6 ms, far from the other IMUs. MVIS estimates it per IMU; it is not a pure time shift (it depends on the motion's frequency), so read it as a compromise value.
- The results files (calibration numbers and run configs, no images) are on Hugging Face beside the data, in its results folder dated 2026-11-02 (see the link above).