rerun/arkitscenes-rrd

ARKitScenes → Rerun (.rrd)

5,015 ARKitScenes indoor iPhone/iPad captures, converted into layered Rerun recordings — including the data that exists only inside the dataset's .mov containers and appears in no published asset:

• 60 Hz camera poses (ARKit visionTransform, 6× denser than the published 10 Hz trajectory) — validated against the published trajectory per sequence and used when a rigid fit agrees within 3° / 10 cm (pose_source = mebx_stream_4_vision_transform), otherwise the published 10 Hz trajectory is slerp-resampled onto the frame grid (pose_source = lowres_traj_slerp)
• 100 Hz raw IMU (accelerometer, gyroscope, fused attitude) — published nowhere else
• Per-frame 8-coefficient lens distortion and wide↔ultrawide stereo extrinsics (~1.2 cm baseline)
• Per-camera clock recovery (the two cameras' PTS origins differ by up to ~100 ms)
• Orientation measured from gravity (the metadata sky_direction label is wrong for ~60% of sampled sequences; the label is kept as a property)
• Full-rate 60 fps video as AV1 VideoStream (av1_nvenc -preset p7 -rc vbr -cq 30, SSIM ≈ 0.984 vs source), ARKit LiDAR depth + laser GT depth + confidence, annotated room mesh and oriented 3D boxes

Structure

Layer-first: one directory per layer, one file per sequence. This makes bulk catalog registration a single register_prefix call per layer, and lets any layer be fetched, regenerated, or re-registered without touching the rest.

base/<video_id>.rrd            # recording properties (queryable columns) + embedded blueprint
calibration/<video_id>.rrd     # 60 Hz world_T_rig, pinholes, lens distortion, stereo extrinsics
video_wide/<video_id>.rrd      # 1920×1440 @ 60 fps AV1
video_ultrawide/<video_id>.rrd
depth/<video_id>.rrd           # ARKit LiDAR depth + laser GT depth + confidence
imu/<video_id>.rrd             # 100 Hz accel / gyro / attitude
gt/<video_id>.rrd              # annotated mesh + oriented 3D boxes
blueprints/                    # dataset-level viewer layouts (portrait default, landscape)

The seven layers share one recording_id per sequence: a Rerun catalog stacks them into a single recording. GT boxes live at stable slot paths (/world/gt/boxes/box_00 …) shared across all sequences — Apple's per-object annotation uid rides along as a uid component on each box — so the dataset's merged schema stays narrow no matter how many sequences are registered. Recording properties (video_id, visit_id, split, pose_source, pose-fit residuals, clock offsets and dispersion, orientation source and quarter-turns, encoder provenance, …) surface as queryable segment-table columns.

Use

# one full sequence (all seven layers)
hf download pablovela5620/arkitscenes-rrd --repo-type dataset --include "*/40753679.rrd" --local-dir .

# one layer for every sequence (e.g. all ground truth)
hf download pablovela5620/arkitscenes-rrd --repo-type dataset --include "gt/*" --local-dir .

# view a sequence (opens with its embedded blueprint)
rerun */40753679.rrd

# or serve a catalog and register everything (rerun >= 0.34)
rerun server --port 51234
# then one register_prefix per layer directory — see the conversion pipeline below

Conversion pipeline (download → mebx decode → clock recovery → ingest → verify): packages/arkitscenes-download — runs locally or fans out on Modal (this dataset was produced on Modal L4 workers; per-recording encoder provenance is embedded).

License & attribution

The underlying data is Apple's ARKitScenes, redistributed with modifications (format conversion to RRD) under the ARKitScenes license — refer to the license text for the exact terms of use. If you use ARKitScenes, cite:

@inproceedings{baruch2021arkitscenes,
  title={{ARK}itScenes - A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile {RGB}-D Data},
  author={Gilad Baruch and Zhuoyuan Chen and Afshin Dehghan and Tal Dimry and Yuri Feigin and Peter Fu and Thomas Gebauer and Brandon Joffe and Daniel Kurz and Arik Schwartz and Elad Shulman},
  booktitle={NeurIPS Datasets and Benchmarks},
  year={2021}
}

Of the 5,047 upstream captures, 24 ship without mesh/annotation/trajectory and 8 failed the pipeline's transcode-integrity gates; both sets are excluded.