Dataset Viewer
Auto-converted to Parquet Duplicate
sequence
stringlengths
5
5
scene
stringlengths
29
53
reference_duration_s
float64
210
561
reference_path_m
float64
111
354
android_frames
float64
6.38k
16.4k
βŒ€
android_duration_s
float64
213
547
βŒ€
android_frames_blurred
float64
381
4.67k
βŒ€
iphone_frames
int64
2.42k
7.63k
iphone_duration_s
float64
120
269
iphone_frames_blurred
int64
63
3.38k
arkit_vs_reference_rms_m
float64
0.1
0.44
flags
stringclasses
9 values
vio01
indoor: corridors of a building
215.7
128.5
6,622
220.56
481
2,424
120.21
188
0.151
android_frames_nearly_black_167_176s iphone_about_20hz iphone_covers_only_the_start
vio02
indoor: corridors and a stairwell over several floors
391.8
249.2
11,832
394.15
381
2,513
129.64
63
0.173
android_dim android_end_trimmed_finger reference_closure_residual_up_to_1.8m multi_floor iphone_about_20hz iphone_covers_only_the_start
vio03
outdoor: paved courtyard with bicycle racks
237.9
180.1
null
null
null
6,744
238.06
987
0.231
no_android
vio04
outdoor: courtyard with planters and steps
313.2
222.7
9,479
315.73
1,168
3,571
120.36
689
0.139
reference_has_no_long_range_closure iphone_covers_only_the_start
vio05
outdoor: road around a block, parked cars
285
282.8
8,576
285.65
4,669
6,153
237.83
3,384
0.444
iphone_image_rotation_check_at_4deg iphone_covers_only_the_start
vio06
outdoor: park with playground equipment
390.5
294.2
11,892
396.11
3,379
3,262
119.97
1,098
0.142
iphone_covers_only_the_start
vio07
outdoor: walled oval field, grass and bare ground
248.1
209.7
7,525
250.64
586
3,013
120.34
470
0.099
iphone_covers_only_the_start
vio08
outdoor: sports ground, running track and pitch
370
353.9
11,222
373.79
3,412
3,536
119.91
787
0.193
reference_has_no_long_range_closure iphone_image_rotation_check_at_4deg iphone_covers_only_the_start iphone_rotation_off_the_rig_value_by_3.3deg
vio11
outdoor: playground on gravel
209.5
110.7
6,380
212.5
3,396
6,259
214.57
3,190
0.247
null
vio12
outdoor at night: an enclosed courtyard under lamps
560.7
285.3
16,407
546.51
1,765
3,578
119.96
479
0.161
night android_smears_when_turning android_end_trimmed_finger iphone_covers_only_the_start iphone_rotation_off_the_rig_value_by_2.63deg
vio13
outdoor at night: the same courtyard, another route
318.4
188.2
9,702
323.16
1,530
7,627
268.89
1,885
0.215
night android_smears_when_turning iphone_covers_only_the_start iphone_rotation_off_the_rig_value_by_3.08deg

Accord Phone Rig Dataset

Code Β· Project page Β· Paper

The 11 walks, one frame of each, and the rig

11 walks on a university campus, recorded by three phones carried together by hand: a Samsung Galaxy Note 9, an iPhone with LiDAR, and a second LiDAR iPhone running RTAB-Map, whose loop-closed trajectory and map are the reference. The phones are not calibrated on a target. It is the phone benchmark of the paper "Accord: Universal Online Metric Localization and Dense Mapping from one Camera and an IMU".

Walks 11: 2 indoors, 7 outdoors by day, 2 outdoors at night
Length 59 minutes, 2.5 km
Recordings 21 (10 Note 9, 11 iPhone)
Frames 99,637 Note 9 (640x480, 30 Hz, IMU 210 Hz); 48,680 iPhone (640x480, about 20 to 30 Hz, IMU 100 Hz, LiDAR depth 256x192)
Reference RTAB-Map on the LiDAR iPhone: a trajectory at about 1 Hz and a coloured map per walk
Size 16.6 GB in 43 archives, 245,671 images
Licence Creative Commons Attribution 4.0 International License (CC BY 4.0)

People, faces and number plates are blurred in every frame.

A look at the data

Every walk at once: the iPhone's frames, sped up.

One moment on both phones (vio11): the Note 9, the iPhone, and the iPhone's LiDAR depth.

The reference (vio11): RTAB-Map's map from above with its trajectory growing, and the iPhone's frame at that moment.

The same clips are in assets/ as GIFs.

Get it

pip install -U huggingface_hub
hf download BE2R/accord-phone-rig --repo-type dataset --local-dir accord-phone-rig
python3 accord-phone-rig/tools/phone_rig_get.py --unpack accord-phone-rig

Older versions of huggingface_hub name the command huggingface-cli. One walk only:

hf download BE2R/accord-phone-rig tools/phone_rig_get.py --repo-type dataset --local-dir accord-phone-rig
python3 accord-phone-rig/tools/phone_rig_get.py --download accord-phone-rig --walks vio07

The second command gets that walk with the tables, calibration and documents, and unpacks it.

On the Hub every folder of images is one tar archive beside it (data.tar, depth.tar, confidence.tar); every other file is stored as it is, so the tables, IMU files and trajectories can be read here. Unpacking gives the layout below. python3 tools/phone_rig_get.py --verify accord-phone-rig checks every unpacked file against SHA256SUMS.

The dataset viewer shows three small tables: walks (sequences.csv), recordings (recordings.csv) and preview (one frame of each recording).

The walks

walk scene walk, s path, m Note 9, s iPhone, s ARKit vs reference, m
vio01 indoor: corridors of a building 216 128 220.56 120.21 0.151
vio02 indoor: corridors and a stairwell over several floors 392 249 394.15 129.64 0.173
vio03 outdoor: paved courtyard with bicycle racks 238 180 - 238.06 0.231
vio04 outdoor: courtyard with planters and steps 313 223 315.73 120.36 0.139
vio05 outdoor: road around a block, parked cars 285 283 285.65 237.83 0.444
vio06 outdoor: park with playground equipment 390 294 396.11 119.97 0.142
vio07 outdoor: walled oval field, grass and bare ground 248 210 250.64 120.34 0.099
vio08 outdoor: sports ground, running track and pitch 370 354 373.79 119.91 0.193
vio11 outdoor: playground on gravel 210 111 212.5 214.57 0.247
vio12 outdoor at night: an enclosed courtyard under lamps 561 285 546.51 119.96 0.161
vio13 outdoor at night: the same courtyard, another route 318 188 323.16 268.89 0.215

"ARKit vs reference" is the RMS after a rigid fit with no scale, at the reference's nodes over the iPhone's part of the walk: the best available measure of how far the reference can be trusted, though not independent of it (RTAB-Map's odometry on its phone is ARKit too). sequences.csv also carries each walk's flags (android_dim, android_end_trimmed_finger, android_frames_nearly_black_167_176s, android_smears_when_turning, ...): what to know before using it.

Layout

After unpacking:

README.md, CHANGES.md, DATASHEET.md, sequences.csv, release.json, SHA256SUMS
calibration/{android,iphone}/   the rig files (OpenVINS and Kalibr format), as the benchmark used them
calibration/rig.json             the phones against each other, one value and each walk's
calibration/iphone_height.json   how far the iPhone rode below the LiDAR iPhone, per walk
tools/                           the scripts that made this from the raw recordings, and an example (tools/README.md)
vioNN/
  sequence.json                  numbers, flags, clock offsets
  android/mav0/cam0/data/<ns>.jpg, cam0/data.csv, cam0/sensor.yaml, imu0/data.csv, imu0/sensor.yaml   (EuRoC layout)
  android/recorder_metadata.json
  iphone/mav0/...                the same
  iphone/arkit.tum               ARKit's own estimate: not a reference
  iphone/depth/<ns>.png, confidence/<ns>.png   LiDAR depth in mm and ARKit's confidence, 256x192
  reference/map.ply              RTAB-Map's map
  reference/trajectory_lidar_iphone.tum   its loop-closed graph (about 1 Hz) on its own clock
  reference/trajectory_{android,iphone}.tum   the same on that device's clock, in that device's IMU orientation

This repository adds recordings.csv (one row per recording), preview/ (one frame of each recording), archives.json and ARCHIVES.sha256 (the archives and their checksums), assets/ and tools/phone_rig_get.py.

Conventions

IMU rows: timestamp [ns], w_x, w_y, w_z [rad/s], a_x, a_y, a_z [m/s^2], specific force as EuRoC. The Note 9 and the iPhone each keep their own clock, whose zero means nothing; the reference's own file, trajectory_lidar_iphone.tum, is on the LiDAR iPhone's clock (Unix time). sequence.json gives the offsets between them (Note 9 and iPhone from their gyroscopes, correlation about 0.998; the LiDAR iPhone from the poses); trajectory_android.tum and trajectory_iphone.tum are already on their device's clock. TUM rows: t x y z qx qy qz qw. Reference positions are the LiDAR iPhone's, lowered by 13.7 cm for the iPhone (its height below the LiDAR iPhone); the other lever arms are not measured, so a comparison has a floor of the rig's size.

The reference and its limits

RTAB-Map on an iPhone: ARKit odometry, LiDAR, and its own loop closures. It is the best estimate here, not ground truth in the survey sense. On vio04 and vio08 it has no long-range closure, so it is odometry; on vio02 a closure disagrees by up to 1.8 m. Its map reaches only about 3.5 m from the track: a map scored against it should leave out its points beyond that (the paper's benchmark leaves out those more than 3 m from the track and reports their share).

Formats

  • Frames: JPEG, 640x480, named by their time in ns; cam0/data.csv lists them. sensor.yaml holds the rig's camera and IMU as in EuRoC, with the IMU noise marked as assumed.
  • iphone/depth/<ns>.png: the iPhone's LiDAR depth, 16-bit PNG in millimetres, 256x192, one per iPhone frame and named as it (frames_without_depth in sequence.json lists the few with none). iphone/confidence/<ns>.png: ARKit's confidence of each depth pixel, 8-bit, 0 low, 1 medium, 2 high.
  • TUM files: t x y z qx qy qz qw, metres and seconds, the pose of the body in the reference's world, whose z is up (gravity is -z). trajectory_{android,iphone}.tum: that device's IMU orientation, at the position described under Conventions. iphone/arkit.tum: ARKit's own estimate in ARKit's own world (the IMU's orientation, the camera's position): not a reference.
  • reference/map.ply: RTAB-Map's map, binary PLY, x y z (float, metres) and colour, in the frame of the reference's TUM files.
  • sequence.json: per walk, scene; clocks (the offsets in s, and the gyroscopes' correlation); per device the frames, duration, rate, IMU samples, the trim at either end, the blur's counts and the converter's checks of the recording against the rig; reference (nodes, duration, path, map points, loop closures, the longest closure's span, closure residuals, ARKit against it); flags, also in sequences.csv.

Example

python3 tools/vio_example.py vio07 (numpy only) reads a walk, prints its recordings and fits ARKit to the reference, as the table's last column.

To run Accord on it: in the Accord code, make phone_rig RELEASE=<the unpacked folder>, then the walks are the sequences vio01 ... vio13 of the setups rig_android and rig_iphone.

Benchmark

The paper runs eight systems on every recording. Means over each phone's recordings: the map's F-score at 5 and 10 cm against the reference map, the map placed by the system's own trajectory and its points more than 3 m from the reference track left out; and the trajectory's error (ATE, m) after a rigid fit. A trajectory more than 10 m off, or none, has failed: it is left out of the ATE mean and its map counts 0.

System Note 9: F@5 F@10 ATE, m iPhone: F@5 F@10 ATE, m
ORB-SLAM3 – – 0.71 – – 0.88
OKVIS2-X – – – – – 0.34
MASt3R-Fusion 0.04 0.08 1.88 Γ— Γ— –
VGGT-SLAM 2.0 † 0.07 0.14 4.76 0.09 0.16 2.72
ScaRF-SLAM's mapper ‑ 0.14 0.26 – 0.10 0.18 –
FoundationVIO 0.15 0.27 2.10 0.13 0.22 1.57
DAVIO 0.12 0.23 0.94 0.11 0.19 1.05
Accord 0.21 0.36 0.94 0.12 0.21 1.06

Bold: best. –: the system gives no such output, or no recording is left to average. Γ—: every recording failed. † no metric scale: given the reference's. ‑ ScaRF-SLAM's mapper on Accord's poses, so no trajectory of its own. All pipelines score far lower here than on the public benchmarks: the phones are not calibrated.

Privacy and use

Bystanders appear in the frames. People (whole bodies), faces and number plates were detected in every frame, linked over neighbouring frames and blurred; the reference maps were checked the same way. There are no names, no audio and no GPS, and the frames carry no EXIF data. The raw recordings are not released. The dataset must not be used to identify, track or describe the people who appear in it. To ask for footage to be removed, write to jaafar.a.mahmoud@itmo.ru or open a discussion in the Community tab.

All walks were recorded in places open to the public. People and the private details in view (faces and number plates) were blurred before release.

DATASHEET.md says how the data was collected and what it is for; CHANGES.md lists everything that was done to the raw recordings.

Licence and citation

The data is under the Creative Commons Attribution 4.0 International License (CC BY 4.0); the scripts in tools/ are under the licence of the Accord code (GPL-3.0). If you use the dataset, please cite:

@misc{mahmoud2026accord,
  title  = {Accord: Universal Online Metric Localization and Dense Mapping from one Camera and an IMU},
  author = {Mahmoud, Jaafar and Movsesyan, Arthur and Huynh, Cuong and Iumanov, Mikhail and Zagoruyko, Sergey and Kolyubin, Sergey},
  year   = {2026},
  eprint = {2611.25279},
  archivePrefix = {arXiv}
}

Jaafar Mahmoud, Arthur Movsesyan, Cuong Huynh, Mikhail Iumanov, Sergey Zagoruyko, Sergey Kolyubin. Biomechatronics and Energy-Efficient Robotics (BE2R) Lab, ITMO University, Saint Petersburg, Russia. Contact: jaafar.a.mahmoud@itmo.ru

Downloads last month
-

Paper for BE2R/accord-phone-rig