Breaking Bad: A Dataset for Geometric Fracture and Reassembly

NeurIPS 2022 Datasets and Benchmarks Track (Featured Paper Presentation)

Silvia Sellán1,*Yun-Chun Chen1,2,*Ziyi Wu1,2,*Animesh Garg1,2,3Alec Jacobson1,2,4
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: