StructureNet

Bibtex #

@article{mo2019structurenet,
  title={Structurenet: Hierarchical graph networks for 3d shape generation},
  author={Mo, Kaichun and Guerrero, Paul and Yi, Li and Su, Hao and Wonka, Peter
          and Mitra, Niloy and Guibas, Leonidas J},
  journal={arXiv preprint arXiv:1908.00575},
  year={2019}
}

Overview #

• Encoding object parts as hierarchical graph
• Taking into account a part hierarchy
• Allows for interpolation in latent space
• Graph (Graph Neural Network ) variational Autoencoder

• Encoder walks from leafs up to the root to end up in latent space

• Decoder walks from the latent vector down and predicts children

• is used as the message aggregation function

• Edges also carry relationship info

• Nodes produce a set of points based on MLP

Use for segmentation #

• Goal: predict part hierarchy and geometry for real-world scans
  1. Train autoencoder on shape parts
  2. Train encoder only for scanned objects (to predict same latent vectors)