TaxonomyProject/SimulationMetadata

PercepTax — Simulation Metadata

Supporting metadata for the PercepTax benchmark.

This repository holds the object taxonomy, object lists / placement rules, and per-scene 3D annotations used to render and annotate the simulated split of PercepTax. It is a collection of JSON/CSV/PNG files (not a load_dataset-style tabular dataset).

Contents

taxonomy/                     # the physical-property taxonomy
├── taxonomy.json             # full taxonomy tree
├── all_objects.json          # object → property assignments
├── material.csv  shape.csv  color.csv  texture.csv
├── affordance.csv  function.csv  environment.csv  physical_properties.csv

objects/                      # object catalog & placement rules
├── background_objects.json
├── objects_list/             # object lists & class↔instance mappings
└── object_placement/         # placable-on-table / -chair / -floor lists

scenes/                       # per-scene 3D annotations
├── annotations/<SceneName>/<view_id>/
│   ├── scene_annotations_split.json   # camera, background, foreground objects,
│   │                                  #   3D boxes, image_path
│   ├── bbox_visualization_all.png     # 2D box overlay
│   └── bbox_3d_visualization.png      # 3D box overlay
├── filter/                            # filtered scene subsets
└── image_quality_ratings*.json        # automated scene quality scores

scene_annotations_split.json fields

scene_name, view_id, image_path, camera (intrinsics/extrinsics), background, foreground (objects with labels and 3D bounding boxes).

Usage

Browse or download with the Hub API:

from huggingface_hub import snapshot_download
path = snapshot_download("TaxonomyProject/SimulationMetadata", repo_type="dataset")

The rendered scene images for the simulated VQA split are bundled in RyanWW/PercepTaxBench (config sim).

License

CC-BY-4.0.