XXXpilar/multiscan-clean

MultiScan Clean

An analysis-ready repackaging of the MultiScan release: 273 RGB-D scans of 117 real indoor spaces, restated as one flat schema covering object boxes, part articulation, structural layout, semantics, meshes, point clouds, camera trajectories and sampled RGB-D frames.

This is a derived work. It carries no scan data the original release does not already publish; it reorganises the annotations and derives a few products (point clouds, layout, relations) that the release leaves implicit.

Provenance and licence

Source: 3dlg-hcvc/MultiScan on the Hugging Face Hub — a gated repository. Licence: CC BY-NC 4.0, inherited unchanged. Non-commercial use only, attribution required.

@inproceedings{mao2022multiscan,
    author    = {Mao, Yongsen and Zhang, Yiming and Jiang, Hanxiao and Chang, Angel X and Savva, Manolis},
    title     = {MultiScan: Scalable RGBD scanning for 3D environments with articulated objects},
    booktitle = {Advances in Neural Information Processing Systems},
    year      = {2022}
}

Note the published release has 273 scans over 117 scenes, not the 230 / 108 quoted in the paper.

Layout

dataset_manifest.json     provenance, coverage, method notes, known limitations
scans.csv                 one row per scan: device, room type, counts, description
objects.csv               one row per annotated object, with its OBB
parts.csv                 one row per part, with its joints
regions.csv               one row per scan: the derived structural layout
relations.csv             pairwise spatial relations
semantic_classes.csv      the corpus-built lexicon (457 classes)
frame_index.json          index of the decoded frames
scans/<scan_id>/
  scan.json               self-contained: region + objects + parts + relations
  cameras.json            intrinsics + camera_to_world, in the mesh frame
  pointcloud.npz          xyz, rgb, normal, semantic_id, instance_id, part_id
  rgb/frame_NNNNNN.png    20 evenly spaced frames
  depth/frame_NNNNNN.npz  metric depth + intrinsics (+ confidence where published)
  normal/frame_NNNNNN.npz camera-space normals from the depth gradient
  mesh/<scan_id>.ply      annotated mesh: vertex colour + per-face object/part ids
  mesh/textured/          <scan_id>.obj + .mtl + texture atlases (tex*.png)

Coordinates are metres, +Z up, floor on the xy plane — MultiScan's native aligned frame. Frame numbers are indices into the original capture, so an RGB, depth and normal file sharing a number describe the same instant, and that number indexes cameras.json.

Contents

scans / scenes	273 / 117
objects	10,957 (100% with an oriented box, 1,174 articulated)
parts	14,941 (3,049 joints)
regions	273
relations	29,947
semantic classes	457
camera poses	4,187,476
RGB frames	5,460
depth / normal frames	5,440
point cloud	54,600,000 points (85.98% labelled)
meshes	273 annotated .ply + 273 textured .obj

Meshes

Both upstream meshes are included, unchanged, for every scan. scans.csv (mesh_ply, mesh_obj) and each scan.json carry their paths relative to the dataset root:

mesh_ply   scans/scene_00000_00/mesh/scene_00000_00.ply
mesh_obj   scans/scene_00000_00/mesh/textured/scene_00000_00.obj

The .ply is the annotated mesh, with vertex colour and per-face object and part ids; it is about 6 GB over all scans. The textured .obj loads its .mtl and atlases from its own folder and is about 58 GB. Most training needs only the point clouds or the .ply; to skip the textures, pass ignore_patterns=["scans/*/mesh/textured/*"] to snapshot_download.

Known limitations

These are properties of the source data, not of the repackaging. The full list is in dataset_manifest.json.

• RGB-D back-projection is not calibrated here. Projecting depth into the mesh frame through the published poses and intrinsics leaves a median 0.069 m residual to the mesh surface, and all 64 rotation × principal-point × axis-sign conventions tried land within 0.069–0.10 m, so the test does not identify a correct convention. Depth, RGB, normals and poses are each written as published; fusing them into a world point cloud needs calibration work first. The mesh-sampled point cloud does not depend on any of this.
• One scan has no depth. scene_00066_01's .depth.zlib fails to inflate (buffer size must be a multiple of element size) on every retry — an upstream file problem. Its 20 RGB frames are intact; frame_index.json records depth: null for it.
• 20 regions have an unreliable room height. Scans rarely look up, so the ceiling is often not annotated and the scan bounds stand in for it; a few scans annotate the ceiling at or below the floor. regions.csv flags these with height_reliable = False — filter on it. Their floor polygons and areas remain valid, which is why they are flagged rather than dropped.
• Wall segments are fragments. MultiScan annotates the wall patches the scan swept, not complete surfaces, so total_wall_length_m understates the true perimeter.
• The footprint is a convex hull of the floor slabs, so it cannot represent a concave or L-shaped room. poly_source records what it was built from.
• Relations and descriptions are templated from geometry, not authored language. They are not human captions and should not be reported as such.
• Articulation state is observed, not authored. The scan caught each joint where it happened to be, so open_fraction is a measurement.

Verification

The build ships a validator covering 31 invariants — box/AABB agreement, joint ranges, unit normals, point-cloud containment, label resolution, rigid camera transforms, depth plausibility, manifest consistency. The release above passes all 31 with no failures and no skips.