wenjingbian/structure_data

structure_data — unified-schema 3D reconstruction datasets

Three datasets for image-to-3D structure prediction (SFT data for Qwen3-VL-4B-Instruct):

• —v11canon — single Infinigen objects (chairs, tables, cabinets, etc.)
• —scene_v3 — multi-object Infinigen scenes
• —physx_v11canon — PartNet-Mobility articulated objects, rotated to v11 chirality

Layout

v11canon/
  train.jsonl           # 7045 rows
  test_seen.jsonl       # 393 rows  (held-out objects in seen categories)
  test_unseen.jsonl     # 390 rows  (objects in unseen categories)
  images/               # *.png referenced by jsonls
scene_v3/
  train.jsonl
  test.jsonl
  images/
physx_v11canon/
  train.jsonl           # 1818 rows
  test.jsonl            # 202 rows
  images/

JSONL format

Each row:

{
  "image": ["v11canon/images/foo_canonical_view.png",
            "v11canon/images/foo_canonical_seg.png",
            "v11canon/images/foo_view9.png"],
  "conversations": [
    {"from": "human", "value": "<prompt with <image> tokens>"},
    {"from": "gpt",   "value": "<compact JSON: materials + nodes + relations>"}
  ]
}

GT JSON top-level fields:

• —materials: dict of {family, desc} keyed by integer index. family is a closed vocab: Wood|Metal|Plastic|Glass|Fabric|Leather|Ceramic|Stone|Rubber|Other. Materials are deduplicated by (family, desc).
• —nodes: dict keyed by integer index (string). Each entry has parent (int or null), material (int as string, key into materials), centroid, extents, and optional joint {type ∈ {hinge, slide}, axis, position, range}.
• —relations: list of {type ∈ {symmetric, instance}, a, b, plane?}.

Conventions

• —Image 1: rendered view of the object/scene with full materials.
• —Image 2: segmentation map at the SAME camera as Image 1.
• —Images 3..N: extra views with possible articulation (no seg map).
• —Coordinates: unit-cube-normalized (max extent = 1.0). Y-up, +X-front.
• —PhysX OBJs are pre-rotated 90° around Y so +Z-front becomes +X-front.