DavidZhang73's picture
Publish source-preserving breaking-bad-volume-constrained dataset
fffa8b6 verified
|
Raw
History Blame Contribute Delete
3.72 kB
metadata
pretty_name: Breaking Bad Volume-Constrained
configs:
  - config_name: default
    data_files:
      - split: full
        path: data/*.parquet

Breaking Bad Volume-Constrained

61,887 fracture assemblies from 696 source objects, with 265,092 embedded fragment meshes. This release contains the official volume-constrained everyday and artifact subsets. It is separate from the original Breaking Bad version and does not include other or GARF point clouds.

from datasets import load_dataset
objects = load_dataset("AssemblyWorld/breaking-bad-volume-constrained", split="all")
sample = objects[0]
parts = sample["parts"]

Pin revision to a release commit for reproducibility. Replace the Hub ID with a local prepared directory for offline loading. Requires datasets >= 5.0.1.

Contents

One row is one fracture, not one original object. sample_id includes the variant, source object path and fracture directory. object_id, subset, nullable category, variant, fracture_id, fracture_type and parts_ct preserve source identity. Both mode_* and fractured_* results are included.

parts contains piece IDs, decoded OBJ paths, float64 vertices, zero-based faces, empty normals and -1 corner normal indices (the source provides no normals). All fragments retain their common source coordinate frame; no placement pose is required. No additional normalization, centering, rotation, repair or sampling is performed. Physical units and upstream normalization remain unknown.

The official decompression procedure maps source triangles to fragments, merges vertices with its 1e-10 tolerance, resolves duplicate faces by orientation, and removes unreferenced vertices. These are decompression operations, not extra cleaning. Shared compressed meshes, sparse maps and labels remain in the raw archive and are not published as separate assets. source_metadata records source hashes, decoder revision and schema/conversion versions.

Source splits

The stored split is full; all loads every assembly once. source_splits is a list of original train/val memberships inherited from the source object. There is no invented test split. Two source objects have no official membership and keep an empty list. Split entries whose objects are absent produce no rows. Do not randomly split fracture rows: fractures of the same object must stay together. No model-specific fragment-count or volume filtering is applied here. The full release has 43,743 everyday and 18,144 artifact samples, with 190,394 and 74,698 fragments respectively. Fragment counts range from 1 to 49. Validation memberships cover 8,025 everyday and 3,738 artifact samples; filtering to 2-20 fragments yields 7,872 and 3,697. The two unassigned objects contribute 198 samples.

validation = objects.filter(
    lambda splits, count: "val" in splits and 2 <= count <= 20,
    input_columns=["source_splits", "parts_ct"],
)

The official volume-constrained generation uses a minimum fragment volume of 1/40 of object volume and rejection sampling; it is not removal of small parts from an existing assembly.

Source and rights

Unofficial conversion of Breaking Bad by Silvia Sellan, Yun-Chun Chen, Ziyi Wu, Animesh Garg and Alec Jacobson (NeurIPS 2022). Official code revision: 899d19a7b0acf1d080521e125c0185940b1b88c7. Schema version 1. The official data record specifies that source meshes inherit ShapeNet and Thingi10K terms, with per-model restrictions. This project's MIT code license does not relicense dataset assets. See the original data record for terms.