AssemblyWorld/fantastic-breaks

Fantastic Breaks

Unofficial source-preserving Parquet conversion of the official CVPR 2023 v1 release. 150 objects, 300 assembly parts and 150 complete reference meshes. One object per row. All geometry and annotations are embedded; no external files are needed.

from datasets import load_dataset
objects = load_dataset("AssemblyWorld/fantastic-breaks", split="all", writer_batch_size=1)
sample = objects[0]
broken, repair = sample["parts"]
complete = sample["complete_reference"]

Requires datasets >= 5.0.1. Pin the published revision for reproducibility. Use writer_batch_size=1 to bound the first-load Arrow cache writer for these high-resolution objects. Arrow formatting avoids expanding geometry into Python lists. The stored split is full; all loads it. No upstream split files are supplied; source_splits is empty for every object. No training/test split is invented.

Contents and provenance

• object_id retains category/object directory IDs, including leading zeros. category_name and material are null because the ZIP contains no label mapping.
• parts holds the real scanned broken mesh and the synthetic_repair proxy. This is not a collection of two independently scanned fracture fragments.
• complete_reference retains model_c.ply separately from assembly inputs. Do not use the reference as model input in a reference-free assembly protocol.
• Meshes preserve float64 coordinates (exact promotion of source floats), original polygon indices/order, existing normals and uint8 RGBA vertex colors. Empty normals mean the source supplied none. Only broken meshes have source RGBA; repair and complete-reference color arrays are empty. ply_header preserves source property declarations and comments; source paths and SHA-256 hashes identify every file.
• annotation retains the original NPZ mask and 4x4 matrix, their dtype strings and shapes. mask indexes the broken mesh's vertices. The matrix is not applied; its direction and physical units are unverified and it is not labeled an SE(3) pose.
• source_metadata records source, archive hash and conversion version. Counts of vertices/faces in local build records include all three meshes; parts counts do not.

Geometry protocol

Conversion does not center, scale, rotate, repair, triangulate, weld, recompute normals or sample geometry. The paper describes upstream scanning, cleaning, alignment and unit-cube normalization. We preserve the PLY coordinates as supplied, without claiming the raw physical scan frame or a verified conversion to meters. The official v1 ZIP contains 150 objects, 450 PLY files and 150 NPZ files. The class/material labels mentioned in the paper are not reconstructed from guesses.

Source and rights

Official project; Google Drive file ID 1pI0UNYkVlveX_ETbdwPXKifUczplZGXn; archive SHA-256 085ba06f1b43841d1f51f868abb37bbc76cfe52bfbef60cd64ff49649eb0f05e (8847570589 bytes). The official archive is available from the original download folder. This is an unofficial storage-format conversion. Source assets retain their original rights; the official v1 archive and project page do not specify a dataset license. The converter's MIT license does not relicense these assets.

Please cite Nikolas Lamb, Cameron Palmer, Benjamin Molloy, Sean Banerjee and Natasha Kholgade Banerjee, Fantastic Breaks: A Dataset of Paired 3D Scans of Real-World Broken Objects and Their Complete Counterparts, CVPR 2023. Paper.

Schema version 1; conversion version 1.