pablovela5620/arkitscenes-mvs-sample

ARKitScenes MVS sample → Rerun (.rrd)

Three ARKitScenes indoor iPhone/iPad captures as layered Rerun recordings. This is the example input of the SimpleRecon / MVS demo in rerun-io/examples-monorepo (packages/mvs): multi-view stereo depth from the video and the camera poses, fused into a TSDF mesh. The ARKit LiDAR depth is included, so you can compare it with the depth that SimpleRecon predicts.

For all 5,015 captures and more layers (ultrawide video, IMU, ground truth), see the full conversion pablovela5620/arkitscenes-rrd.

Captures

id	room	seconds	video frames	LiDAR depth frames	layer files	recording
42897542	living room	30.15	1,810	1,810	119.1 MB	119.1 MB
48458235	kitchen	24.65	1,480	1,480	91.3 MB	91.3 MB
48018972	kid's room	36.94	2,218	2,217	125.2 MB	125.3 MB

Structure

Layer-first: one directory per layer, one file per capture. Every file of a capture uses the capture id as its recording_id, so a Rerun catalog stacks the four layers into one segment. All time-varying data is on the video_time timeline (duration since the first frame). The world is Z-up; cameras are RDF (right, down, forward).

base/<id>.rrd          # recording properties (split, orientation, pose source, pose-fit residual, ...),
                       #   world coordinates (/world), rig and camera frames
calibration/<id>.rrd   # 60 Hz world_T_rig (/world/rig_00), static rig_T_cam (/world/rig_00/cam_00),
                       #   wide camera pinhole 1920x1440 (/world/rig_00/cam_00/pinhole),
                       #   depth-resolution pinhole 256x192 (/world/rig_00/cam_00/pinhole_lowres)
video_wide/<id>.rrd    # wide camera, 1920x1440 @ 60 fps AV1 VideoStream (/world/rig_00/cam_00/pinhole/video)
arkit_depth/<id>.rrd   # ARKit LiDAR depth, 256x192 16-bit PNG in mm (/world/rig_00/cam_00/pinhole_lowres/depth)
                       #   and its confidence (/world/rig_00/cam_00/pinhole_lowres/confidence)
recordings/<id>.rrd    # NOT a layer: the four layers of one capture in one file, with a viewer layout
                       #   (3D view: the moving camera frustum with the video frame and the LiDAR depth as points;
                       #   2D views: the video, the LiDAR depth)

The calibration layer also carries the ultrawide camera's pinhole (/world/rig_00/cam_01); its video is not part of this sample.

Use

# everything (about 670 MB)
hf download pablovela5620/arkitscenes-mvs-sample --repo-type dataset --local-dir arkitscenes-mvs-sample

# only the four layer files of one capture
hf download pablovela5620/arkitscenes-mvs-sample --repo-type dataset --include "*/42897542.rrd" --exclude "recordings/*" --local-dir arkitscenes-mvs-sample

# look at one capture (Rerun viewer 0.42 or newer; opens with its own layout)
rerun arkitscenes-mvs-sample/recordings/42897542.rrd

The video is AV1. Use a Rerun viewer or a browser with AV1 decode (recent Chrome or Firefox, or Safari on hardware with AV1 decode).

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}
}