Commercially licensable real-capture indoor data, focused on hospitality.

Head-to-head against the leading real-capture and synthetic indoor-scene datasets. Numbers are from each project’s own page or peer-reviewed paper, verified 2026.

ScanNet++ delivers sharper geometry (sub-millimeter laser) and HM3D delivers larger buildings; Hypersim and 3D-FRONT deliver perfect per-pixel labels because they are synthetic. Every dataset above is research-only or non-commercial except the synthetic ones, and none of them is built around hotel rooms. Roomza is the only real-capture, commercially-furnished, hospitality-specific corpus available under a commercial license.

Hotel rooms are one of the most repeatable indoor environments in the world. Beds, desks, bathrooms, windows, lighting, doors, storage, soft goods, reflective surfaces, tight paths, and accessibility variations, across thousands of real-world layouts, finishes, and price points. Repeatable enough to be a useful training distribution, varied enough to generalize from. They are also one of the few indoor categories where capture rights and commercial licensing line up cleanly.

One row per asset type. Label inventory is listed honestly: raw, labeled, or routed through the labeling track. Primary formats follow ScanNet and ARKitScenes conventions; COCO ships as a 2D export, not as the canonical container.

Phone capture is the scaling tier, not a substitute for survey-grade rigs. Geometric accuracy is named on every capture so a buyer can scope coverage to their tolerance.

What ships today, where, and what is on the roadmap. We’d rather scope a license to what fits than oversell coverage we don’t have.

A real record from the sample bundle. IDs are stable across schema versions, captures, and label refreshes. Sibling files (stills, video manifest, IMU CSV, ScanNet-style instance aggregation, COCO export) carry the same room_id.

{
  "room_id":        "rm_corner_king_1108",
  "schema_version": "1.0.0",
  "archetype":      "boutique_king",
  "property_class": "boutique",
  "corner_room":    true,
  "dimensions_mm": {
    "width_x": "…", "length_z": "…", "ceiling_height_y": "…"
  },
  "bed_config": {
    "primary_bed": "king",
    "mattress_size_mm": { "w": "…", "l": "…", "h": "…" }
  },
  "accessibility": {
    "ada_compliant": false,
    "roll_in_shower": false,
    "visual_alarm": true,
    "hearing_kit_available": true
  },
  "capture": {
    "device_rig": "rig_iphone15pro_polycam_v3",
    "labeled":    true
  },
  "hotel_id_anonymized": "hp_xxxxxx",
  "city_anonymized":     "us_pnw_metro_a"
}

What's the primary data format?

Per-room structured JSON record plus sibling artifacts (stills, video manifest with sidecar IMU CSV, RoomPlan USDZ where captured, ScanNet-style instance aggregation JSON where labeled). 2D detection ships as a COCO 1.0 export when requested; ARKitScenes-format oriented 3D bboxes ship in the instance aggregation manifest.

What taxonomy do labels use?

A Roomza taxonomy designed for furnished commercial lodging interiors, with explicit crosswalks to NYUv2-40, ScanNet-20, and ARKitScenes-17. A meaningful share of the class set is distinctive to hospitality and has no equivalent in the public reference taxonomies. The full taxonomy is browsable from the sample bundle.

Are camera intrinsics provided?

Yes. Camera intrinsics are acquired directly from ARKit at capture time, not estimated from imagery. They ship in the per-video manifest for walkthroughs and on RoomPlan spatial captures, with EXIF retained on imported stills. Camera pose (extrinsics) is a separate channel, also from ARKit at capture and handled independently of intrinsics.

How accurate is the geometry?

Honest tiering, named on every capture. Phone RGB walkthroughs are monocular RGB with ARKit camera pose captured per frame. Phone-LiDAR RoomPlan captures land at consumer-LiDAR wall accuracy. Survey-grade reference captures (Matterport Pro3-class) deliver survey-grade point clouds and panoramic RGB-D.

What's the per-room data volume?

Typical: one structured record, multiple stills, optional walkthrough video, optional spatial capture, plus instance aggregation and 2D label exports on the labeled subset. The sample bundle shows the shape on coherent example rooms.

Why a labeling track instead of fully-labeled everything?

Labeling 6,000+ properties exhaustively before licensees know what they need is wasteful. Scope the subset, the class set, and the label spec; we route through vetted partners under license and ship the bundle. Pricing is per-image or per-frame and quotes are written.

When does LiDAR or ground-truth depth reach network scale?

Expanding via partner-device coverage. Reliable today on the iOS-equipped capture subset. Talk to us if you need depth at scoped scale and we will model coverage to your need.

Why is accessibility metadata, not a perception task on this corpus?

ADA-relevant fields are powerful as a corpus-filter dimension and as a structured signal. Building an accessibility robot is still a perception + navigation problem against sensor data; the metadata isn't a substitute for that. We surface accessibility as a filter, not as a task category.