Reproduce “Category-Level Multi-Part Multi-Joint 3D Shape Assembly” 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
Yichen Li, Kaichun Mo, Yueqi Duan, He Wang, Jiequan Zhang, Lin Shao, Wojciech Matusik, Leonidas Guibas
This work develops a hierarchical graph learning approach for category-level multi-part multi-joint 3D shape assembly, precisely predicting 6-DoF poses and aligning peg-hole joints. The method achieves an average Part Pose Accuracy of 69.2% and Joint Matching Accuracy of 51.4%, effectively generating physically assemblable structures where baselines often failed to produce valid shapes.
Key insights
- Explicitly modeling functional joints and their precise alignment is crucial for producing physically viable 3D shape assemblies, moving beyond purely abstract geometric considerations.
- A hierarchical, interleaved optimization strategy that first establishes global shape structure and then refines it with local joint information effectively addresses the intertwined challenges of discrete matching and continuous pose prediction.
- The use of specialized loss components and a custom order-invariant joint matching algorithm is essential for handling rotational and reflective symmetry ambiguities in parts and ensuring consistent supervision during training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.