Breaking Bad Dataset

Breaking Bad: A Dataset for Geometric Fracture and Reassembly

NeurIPS 2022 Datasets and Benchmarks Track (Featured Paper Presentation)

Silvia Sellán^1,*, Yun-Chun Chen^1,2,*, Ziyi Wu^1,2,*, Animesh Garg^1,2,3, Alec Jacobson^1,2,4
¹University of Toronto ²Vector Institute ³NVIDIA ⁴Adobe Research, Toronto
^* Equal contribution

[Paper] [Review] [Data] [Benchmark Code] [Data Generation Code]

[Twitter Thread] [Alternative Data Host (Kaggle)]

Dataset

Our Breaking Bad Dataset contains around 10k meshes from PartNet and Thingi10k. For each mesh, we pre-compute 20 fracture modes and then simulate 80 fractures from them, resulting in a total of 1,047,400 breakdown patterns. We divide our dataset into three subsets, namely, everyday, artifact and other to faciliate different applications.

Accessing the dataset

See README for detailed instructions.

Benchmark

Please see our benchmark results on geometric assembly in the paper. We release our benchmark code at Github.

Gallery

Our dataset is divided into three categories:

The artifact category consists of fractures of meshes from Thingi10k which are tagged as "sculpture" or "scan", for use in archelogical reconstruction. Click here for a gallery view of some results.
The everyday category consists of fractures of all the meshes from the following 20 categories of the PartNet dataset (click on the names for a result gallery view): BeerBottle, Bottle, Bowl, Cookie, Cup, DrinkBottle, DrinkingUtensil, Mirror, Mug, PillBottle, Plate, Ring, Spoon, Statue, Teacup, Teapot, ToyFigure, Vase, WineBottle, WineGlass.
The other category contains fractures of all other meshes in Thingi10k. Click here for a gallery view of some results.