Reproduce “PartMat: Material-Aware 3D Part Decomposition with a Single Global Latent” 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

Guangming Fu, Jin Song, Yiyun Fei, Guoqiu Li, Ruigao Yang, Jianan Jiang

PartMat introduces a three-stage pipeline for material-aware 3D part decomposition from an image and whole-object mesh, representing multi-part geometry with a single global latent vector. The method achieves competitive geometric reconstruction with a Chamfer Distance of 2.83x10⁻⁴ and an F1-Score of 69.59, while enabling efficient, editable material assignments, significantly outperforming prior methods in inference efficiency for multi-component objects.

Key insights

  • Representing multi-part geometry with a single global latent vector in PartVAE enables constant decoding cost irrespective of the number of material components, addressing scalability limitations.
  • Direct-gradient reinforcement learning alignment with a differentiable SDF reward (match and overlap penalties) effectively enhances material boundary precision and consistency during latent generation.
  • The combination of a multi-channel implicit SDF decoder and a sparse, part-aware geometry refinement stage facilitates both efficient multi-part representation and the recovery of fine-grained geometric details.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.