Splataverse-Replica: Replica Dataset with real per-Gaussian semantic labels

Every scene in facebookresearch/Replica-Dataset v1.0 (18 real scanned indoor spaces), converted through splataverse's own analytic, training-free mesh->3D-Gaussian-splat converter, with a REAL per-Gaussian seg_id carried the entire way from Replica's own habitat/info_semantic.json per-instance class_id annotation -- not a hand-assigned or predicted label.

License

Non-commercial research/educational use only, per the Replica Dataset's own "Replica Dataset Research Terms" (facebookresearch/Replica-Dataset LICENSE): employees of for-profit/commercial entities may use it only for research/educational purposes, not commercial products; provided AS IS with no warranty; California law governs. This derivative (splat conversion + real semantic labels) carries the same terms as the source data.

Structure

Per scene (<scene_name>/):

• scene.ply -- one merged composite splat for the whole scene. Standard splat PLY fields (x,y,z,scale_0..2,rot_0..3,opacity,f_dc_0..2) plus a custom int32 seg_id property: Replica's own real semantic class_id for the object each Gaussian came from (-1 for objects absent from Replica's own annotation, or explicitly labeled "undefined" there -- a real gap in the source dataset's own export, not this pipeline's).
• manifest.json -- per-object category name, class_id, Replica's own instance id, and Gaussian count.

classes.json (repo root) -- the real 101-class id->name map, identical across every scene (Replica uses one shared taxonomy for the whole dataset).

Provenance

Converted via splataverse.convert_mesh (unmodified converter, same code path every other object in this project's datasets uses) on habitat/mesh_semantic.ply + info_semantic.json, grouping faces by Replica's own per-face object_id, fan-triangulating each object's quads (this mesh format), converting per-object, then stamping every resulting Gaussian's seg_id with that object's real class_id.