Reproduce “PartSAM: A Scalable Promptable Part Segmentation Model Trained on Native 3D Data” 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

Zhe Zhu, Le Wan, Rui Xu, Yiheng Zhang, Honghua Chen, Zhiyang Dou, Cheng Lin, Yuan Liu and 1 more

PartSAM is introduced as the first scalable, feed-forward, and promptable 3D part segmentation model, leveraging native 3D data and a dual-branch architecture to segment both visible and occluded object structures. It achieves a 91% relative improvement in interactive segmentation over Point-SAM on PartObjaverse-Tiny and consistently surpasses state-of-the-art methods in automatic class-agnostic segmentation.

Key insights

  • Leveraging a dual-branch encoder that combines frozen 2D priors with learnable 3D semantics from native data is effective for robust 3D part understanding and generalization.
  • Training on a large-scale, high-quality dataset of native 3D part annotations is essential to overcome surface-only understanding and enable the segmentation of intrinsic 3D geometry.
  • A prompt-guided, feed-forward architecture allows for flexible user interaction and automatic decomposition, mirroring the promptability of 2D foundation models.

Summary by alphaXiv.

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