Reproduce “Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bi'an Du, Xiang Gao, Wei Hu, Renjie Liao
A deep learning model for generative 3D part assembly leverages an explicit part-whole hierarchy, inferring super-part poses to guide the placement of individual components. The proposed method achieves state-of-the-art performance on the PartNet dataset, demonstrating up to a 2.99% increase in mean Part Accuracy and a 2.95% increase in mean Connectivity Accuracy for chairs, alongside superior visual quality confirmed by human studies.
Key insights
- Explicitly modeling part-whole hierarchies allows super-part poses to provide strong, simplifying hints for the poses of their constituent smaller parts.
- Predicting the poses of a fewer number of super-parts is an inherently easier task than simultaneously predicting all individual part poses, simplifying the overall assembly problem.
- Latent super-part pose prediction, learned solely through backpropagation, eliminates the need for costly ground-truth supervision for these higher-level components.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.