Reproduce “NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction” 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
Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, Lingjie Liu
A method for rapidly learning neural implicit surfaces from multi-view images combines high-quality 3D geometry with significantly accelerated training times. It reconstructs static scenes in 5 minutes and processes dynamic video frames in as little as 20 seconds per frame, outperforming prior methods in speed and geometric accuracy.
Key insights
- Combining multi-resolution hash encodings with an SDF-based volume rendering framework allows for both accelerated training and high-fidelity surface reconstruction.
- Efficient and precise computation of second-order derivatives, particularly with ReLU-based MLPs, is crucial for stable and fast training of neural implicit surfaces that rely on Eikonal regularization.
- Leveraging temporal coherence through incremental training and a global transformation prediction component enables robust and rapid neural surface reconstruction for complex dynamic scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.