Splataverse-Hoi: articulated indoor objects at multiple openness states, real 3D Gaussian splats

Real Leica laser-scanned locations from the Hoi! dataset (Engelbracht et al., CVPR 2026, arXiv:2512.04884), converted through splataverse's own analytic, training-free mesh->3D-Gaussian-splat converter. Every location has 2+ Leica scans ("setups") of the same physical space at different times -- i.e. different real articulation states of whatever drawers/doors/cabinets are in it (a door open vs. closed, a drawer pulled out vs. pushed in).

The task this is built for

Given two (or more) real Gaussian-splat reconstructions of the same articulated part at different real states, predict how it articulates: its joint type ("revolute" or "prismatic"), its axis (a 3D unit direction), and -- for revolute joints -- the position of a point on that axis. A real 3D handle_point_3d (where a human actually grasped the part, lifted from the dataset's own manual annotation) is provided as a physically-informative input alongside the splats, not something to predict.

Everything needed to attempt this for one object lives in that object's own folder -- <location>/objects/object_<NNN>/:

object_001/
  state_001/scene.ply   # one real captured state (Gaussian splat, real color)
  state_002/scene.ply   # a different real state of the SAME part
  label.json            # {type, axis, position, handle_point_3d, description, ...}

No need to cross-reference the location's full manifest -- label.json is self-contained. See label.json's own type_conflict_with_instance_annotation field before trusting type blindly for a handful of parts (see the data-quality note below) -- everything else in label.json is unconditionally real.

License

CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) -- commercial use permitted with attribution. Cite:

@InProceedings{Engelbracht_2026_CVPR,
    author    = {Engelbracht, Tim and Zurbr\"ugg, Ren\'e and Wohlrapp, Matteo and B\"uchner, Martin and Valada, Abhinav and Pollefeys, Marc and Blum, Hermann and Bauer, Zuria},
    title     = {Hoi! - A Multimodal Dataset for Force-Grounded, Cross-View Articulated Manipulation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June}, year = {2026}, pages = {8880-8890}
}

Structure

Per location (e.g. oven_1/):

• objects/object_<NNN>/ -- the primary data for the articulation task (see above): one state_<NNN>/scene.ply per real captured state of that one articulated part, tightly cropped (no whole-room background), plus a self-contained label.json. State 1's crop comes directly from the dataset's own instance-segmentation annotations (manual 2D panorama masks lifted to 3D); other states have no per-state instance annotation, so we crop that state's own point cloud using state 1's instance bounding box, extended generously along the joint's own known sliding axis so the object is still captured wherever it ended up.
• whole_room/state_<NNN>/scene.ply -- the full room scan for that Leica setup, supplementary context (not needed for the per-object task above). Real color throughout, uncapped Gaussian budget. State 1 starts from the dataset's own pre-built mesh (Leica's proprietary "3DReshaper" scanner software output, shipped with no color/UV); every other state has no shipped mesh, only a raw point cloud, so we build our own mesh via Open3D Poisson reconstruction + a density-based trim. Known limitation, found and traced this session: some locations' whole-room reconstruction is genuinely rough -- confirmed (by inspecting the raw Leica point cloud itself, before any of our processing) to be real scan-registration artifacts in the source capture, not something our meshing introduces or that reconstruction-parameter tuning can fix. The per-object crops above are unaffected -- verified directly on the same locations.
• manifest.json -- the full per-location record: every object's metadata (including a resolved type field, identical to what each label.json carries) plus both of two disagreeing articulation-type labels the dataset itself ships for the same parts (<location>.json's own type field vs. the instance annotation's class field -- found to genuinely conflict for oven_1's 9 parts: prismatic vs. revolute). type uses the joint-file's label, since axis/ position values are only meaningful evaluated against the type they were annotated under; the disagreement is flagged, never silently hidden.
• contact_sheet.png (currently oven_1 only) -- one real gsplat render per whole-room/object x state cell, for a quick visual sanity check of what's actually in the data.

Provenance

Source data: bonndata (University of Bonn Dataverse), DOI 10.60507/FK2/QODWTV. Converted via splataverse.convert_mesh (unmodified converter, same code path every other object in this project's datasets uses), uncapped Gaussian budget per explicit request.

Status: all 20 released locations converted -- oven_1, oven_2, bathroom_1, bathroom_2, bedroom_1, bedroom_4, bedroom_5, bedroom_6, bedroom_8, bedroom_10, bedroom_11, fridge_1, kitchen_7, kitchen_14, kitchen_15, kitchen_17, livingroom_1, office_1, wardrobe_1, wardrobe_2 -- 195 real articulated parts, 183 with a resolvable ground-truth label (the remainder have no entry in the source <location>.json joint file).

A companion challenge/leaderboard around this data is planned but not yet hosted.