Reproduce “Neural Points: Point Cloud Representation with Neural Fields for Arbitrary Upsampling” 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
Wanquan Feng, Jin Li, Hongrui Cai, Xiaonan Luo, Juyong Zhang
Researchers from the University of Science and Technology of China developed "Neural Points," a point cloud representation that encodes local continuous geometric shapes via neural fields, enabling arbitrary-factored upsampling. This approach consistently outperforms existing methods on Chamfer Distance, Hausdorff Distance, and Point-to-Surface metrics, generating smooth, high-fidelity surfaces while demonstrating robustness to noise and generalization to unseen data.
Key insights
- Representing each point as a local continuous neural field allows for a more information-rich, resolution-free definition of a "point" in a point cloud.
- Integrating multiple local neural fields through weighted projection creates a globally consistent and continuous surface, combining the benefits of local detail and global coherence.
- Effective supervision of a continuous surface from discrete ground truth is achieved using projection-based Chamfer Distance, normal consistency loss, and an integration loss for seamless transitions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.