Reproduce “Hierarchical Neural Semantic Representation for 3D Semantic Correspondence” 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

Keyu Du, Jingyu Hu, Haipeng Li, Hao Xu, Haibing Huang, Chi-Wing Fu, Shuaicheng Liu

A training-free framework called Hierarchical Neural Semantic Representation (HNSR) is presented for robust 3D semantic correspondence, leveraging implicit priors from pre-trained 3D generative models. This approach, which includes a progressive global-to-local matching strategy, achieves accurate and generalizable semantic mappings for tasks like co-segmentation and keypoint matching across diverse shapes.

Key insights

  • Pre-trained 3D generative models implicitly encode valuable geometric and semantic priors from vast datasets, which can be directly extracted and leveraged for 3D correspondence without further training.
  • A hierarchical feature representation, combining deep-layer global semantics with shallow-layer local geometry, enables a robust global-to-local matching strategy that ensures both high-level semantic consistency and fine-grained geometric accuracy.
  • Developing a training-free method, by effectively repurposing existing generative models, overcomes the significant burden of data annotation and enhances the scalability and generalization capabilities of 3D correspondence.

Summary by alphaXiv.

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