Reproduce “SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration” 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
Xu Cao, Takafumi Taketomi
SuperNormal presents a 3D reconstruction approach by integrating multi-view surface normal maps with a neural implicit surface representation, yielding high-fidelity geometry with fine details. The method achieves significantly faster training times through a novel directional finite difference technique for computing Signed Distance Function (SDF) gradients.
Key insights
- Integrating high-quality multi-view normal maps into neural implicit surfaces effectively addresses the detail-smoothing limitation of color-image-based neural reconstruction, achieving detail levels comparable to commercial 3D scanners.
- SDF gradient computation can be substantially accelerated by strategically reusing existing SDF samples for directional finite differences, bypassing the inefficiencies of traditional automatic differentiation or finite difference methods.
- The combination of multi-resolution hash encoding for neural implicit fields and supervision from normal maps provides both expressive power and efficiency, leading to robust and detailed 3D reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.