Reproduce “Learning Compact Latent Space for Representing Neural Signed Distance Functions with High-fidelity Geometry Details” 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
Qiang Bai, Bojian Wu, Xi Yang, Zhizhong Han
A novel framework combines generalization-based and overfitting-based strategies with a dual-branch architecture and a shared compact latent code to represent multiple Neural Signed Distance Functions. This approach achieves high-fidelity geometric details while maintaining a compact, generalizable latent space, outperforming existing methods on 3D reconstruction, completion, and interpolation tasks.
Key insights
- The synergy between global generalization and local overfitting branches, linked by a shared latent code, effectively addresses the trade-off between detail fidelity and representational compactness.
- Dynamically fusing predictions based on proximity to the surface allows the system to leverage each branch's strengths for different spatial regions.
- Intelligent, balanced sampling near the surface within a shared grid is crucial for multi-object, high-fidelity detail recovery, preventing inter-shape interference and improving efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.