XXXpilar/hssd_clean

HSSD-200 Clean

An analysis-ready rebuild of HSSD-200 (hssd/hssd-hab). The source release spreads scene layout, object placement, semantics and physics across several files and coordinate conventions; this resolves them into one schema and derives the products the release only implies — oriented bounding boxes, room polygons, labelled point clouds, camera poses, and rendered RGB / depth / normal maps.

This is a derivative work. All geometry, semantics and appearance originate from HSSD-200 by Khanna et al. Please cite them (below) and respect the non-commercial licence.

Contents

File	Rows / count	What it is
objects.csv	56,647	one row per placed object instance
regions.csv	2,351	one row per room
asset_catalog.csv	15,694	one row per unique object model used
semantic_classes.csv	406	the semantic lexicon (id ↔ name)
dataset_manifest.json	—	provenance, coverage, method notes, validation, limitations
render_index.json	—	index of every rendered view
scenes/<scene_id>/scene.json	168	self-contained scene: layout + all objects
scenes/<scene_id>/pointcloud.npz	168	34.6 M points total
scenes/<scene_id>/cameras.json	168	9,404 poses total
scenes/<scene_id>/rgb,depth,normal/	2,016 each	rendered views
meshes/objects/	15,681 .glb	object meshes, including doors/windows under objects/openings/
meshes/stages/	168 .glb	per-scene architecture shell (walls, floors, ceilings)
meshes/urdf/	1,509 models	URDF + per-link .glb for articulated objects

Coordinates are metres, +Y up, −Z front, right-handed — unchanged from Habitat, so results stay comparable with other HSSD work.

Quick start

import pandas as pd, json, numpy as np
from huggingface_hub import snapshot_download

root = snapshot_download("XXXpilar/hssd_clean", repo_type="dataset")

objects = pd.read_csv(f"{root}/objects.csv",
                      dtype={"semantic_id": "Int64", "region_id": "Int64"},
                      converters={"obb_center": json.loads,
                                  "obb_half_extents": json.loads,
                                  "aabb_min": json.loads, "aabb_max": json.loads})

cloud = np.load(f"{root}/scenes/102344280/pointcloud.npz")
xyz, rgb, sem, inst = cloud["xyz"], cloud["rgb"], cloud["semantic_id"], cloud["instance_id"]

List-valued columns are JSON strings in the CSVs and real arrays in the per-scene JSON. The explicit dtype keeps id columns integral — pandas otherwise widens them to float.

Key fields in objects.csv

Boxes. obb_center + obb_half_extents + obb_rotation_wxyz is the tight oriented box. aabb_min / aabb_max is its world-axis-aligned envelope. Half-extents are already scaled — do not re-apply non_uniform_scale.

Use the OBB, not the AABB, when tightness matters. An AABB must bound the box's corners, so under a non-axis-aligned yaw it is legitimately looser than the mesh's own AABB.

Rooms. region_id joins regions.csv on (scene_id, region_id). region_source says how: inside (48,569) centre within the polygon; nearest (3,670) centre inside wall thickness, matched to the nearest room within 1 m — almost entirely doors and windows, which are part of a wall and by construction fall outside every polygon; none mostly loose outdoor props. Filter region_source == 'inside' for strict containment.

Semantics. semantic_id is derived from category via the lexicon, so the two are always consistent. semantic_id_habitat holds the asset's object_config id separately, because that value is sometimes wrong (spot checks found a "Book" labelled animal, a "Pillow" labelled awning); it is promoted only when the category has no lexicon entry.

Articulation. Rows with instance_kind == 'articulated' (3,626) carry joint_states — per joint: name, type, position, limits, open_fraction — plus n_movable_joints, base_type and urdf_path. 10,607 joints total (5,115 revolute, 5,482 prismatic).

Meshes. render_asset, urdf_path and the scene's stage_render_asset are repo-relative paths that resolve under meshes/, which mirrors the hssd-hab layout:

import trimesh
row = objects.iloc[0]
mesh = trimesh.load(f"{root}/meshes/{row.render_asset}")      # asset-local frame
stage = trimesh.load(f"{root}/meshes/stages/102344280.glb")    # already in world frame

Object meshes are in the asset's local frame; place them with p_world = translation + R(obb_rotation_wxyz) · (non_uniform_scale ⊙ p_local). render_asset_present is True for 56,619 of 56,647 rows; the rest are the dangling templates below and a few articulated models whose geometry exists only under urdf/.

How the boxes were derived

Placements come from scenes/*.scene_instance.json; extents from the mesh where available and from aligned.dims otherwise. Rigid placements use translation_origin: "asset_local":

p_world = translation + R(rotation) · (non_uniform_scale ⊙ p_local)

Three conventions are undocumented in the source and were established empirically:

1. Quaternion order is [w,x,y,z] — under that reading 76% of the 53,021 placements are pure yaw and 8,911 are exactly identity; under [x,y,z,w], zero are.
2. aligned.dims is in the canonical frame but the .glb is not — for the 2,895 assets declaring a non-default up/front, dims are rotated into the mesh frame first, otherwise their axes come out permuted.
3. Assets sit base-at-origin, x/z centred (metadata path only).

Validation

Mesh vertices projected into each box's own frame across sampled instances: the OBB is tight to 1.5e-06 m — nothing escapes it, no measurable slack. Every sampled object's point cloud falls inside its own recorded AABB. 35.6% of placed objects rest within 2 cm of their room floor and 0.51% below it. All rendered normals are unit-length on surfaces and exactly zero on background.

Limitations

RGB colour is partial. Many HSSD materials store colour in KTX2/Basis textures (KHR_texture_basisu) that available Python decoders cannot read. Across all 34.6 M points:

Source	Share
real texture	27.6%
baseColorFactor (linear→sRGB)	35.2%
white placeholder (KTX2-only material)	37.2%

Geometry, depth, normals, semantics and instances are unaffected.

Renders are pyrender, not habitat-sim, so RGB is not photometrically comparable to habitat renders. Depth and normals are geometric and unaffected — normals come from a flat normal-encoded pass, not from differentiating depth.

Camera poses are synthesised, not captured trajectories: a ring of views about each room centroid at 1.5 m eye height.

Articulated objects have no bounding box — they use translation_origin: "COM" with per-link URDF geometry, so joint state is recorded rather than a box guessed. Box coverage is therefore 99.97% of rigid instances and 93.57% of all rows.

Scale. The public hssd-hab release contains 168 scenes and 16,541 shipped object models, not the 211 scenes / 18,656 models quoted for HSSD-200 in the paper. 15,694 models actually appear in the scenes.

A few dangling references. Templates 3759, 3760 and 5804 are referenced by scenes but absent from the source repo. (5849 was once counted here too; it lives in objects/openings/ and its mesh is now included.)

Door and window boxes are still estimated. Door/window meshes live in objects/openings/, which earlier builds did not look in, so their boxes (2,956 instances, bbox_source == 'metadata') come from aligned.dims under the base-at-origin assumption, and they are missing from the point clouds. The meshes are now in meshes/; a rebuild will replace these with measured boxes.

Licence and attribution

Released under CC-BY-NC-4.0, matching the source. Non-commercial use only, attribution required. Files under meshes/ are unmodified copies from hssd/hssd-hab and remain subject to its terms. If you use this, cite the original HSSD work:

@inproceedings{khanna2024hssd,
  title     = {{Habitat Synthetic Scenes Dataset (HSSD-200)}: An Analysis of 3D Scene
               Scale and Realism Tradeoffs for ObjectGoal Navigation},
  author    = {Khanna, Mukul and Mao, Yongsen and Jiang, Hanxiao and Haresh, Sanjay and
               Shacklett, Brennan and Batra, Dhruv and Clegg, Alexander and
               Undersander, Eric and Chang, Angel X. and Savva, Manolis},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and
               Pattern Recognition (CVPR)},
  year      = {2024}
}

Source dataset: https://huggingface.co/datasets/hssd/hssd-hab · Project page: https://3dlg-hcvc.github.io/hssd/