# NeRFMeshing: Distilling Neural Radiance Fields into Geometrically-Accurate 3D Meshes

**Authors:** Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, Federico Tombari  
**arXiv:** [2303.09431](https://arxiv.org/abs/2303.09431)  
**Date:** 2023-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.09431v1/

## Key points

- Neural Radiance Fields (NeRFs) revolutionized view synthesis by proposing that each 3D point can emit radiance
- NeRFs enable differentiable volumetric rendering for accurate representation of 3D scenes
- Traditional 3D meshes are still primary in computer graphics due to real-time rendering and physics-based simulations support
- Extracting 3D meshes from NeRFs is challenging as they prioritize view synthesis over precise geometry enforcement
- Researchers proposed NeRFMeshing as a solution for easy 3D surface reconstruction from any NeRF-driven model
- NeRFMeshing uses a compact architecture to distill the volumetric representation into a Signed Surface Approximation Network
- The resulting 3D mesh from NeRFMeshing is physically accurate and supports real-time rendering on various devices
- NeRFMeshing bridges the gap between neural radiance fields and traditional 3D mesh representations, enhancing visualization and simulation capabilities

## Summary

Neural Radiance Fields (NeRFs) have revolutionized the field of view synthesis by proposing that each 3D point can emit radiance. This enables differentiable volumetric rendering for accurate representation of 3D scenes. While NeRFs excel at image rendering, traditional 3D meshes remain the primary scene representation in computer graphics and simulation pipelines due to their support for real-time rendering and physics-based simulations. However, extracting 3D meshes from neural radiance fields poses a significant challenge as NeRFs prioritize view synthesis over enforcing precise underlying geometry on the radiance field. In response to this challenge, a team of researchers including Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari have proposed a novel solution called NeRFMeshing. This approach introduces a compact and flexible architecture that facilitates easy 3D surface reconstruction from any NeRF-driven model. By training the radiance field and distilling the volumetric 3D representation into a Signed Surface Approximation Network, NeRFMeshing allows for seamless extraction of geometrically-accurate 3D meshes and appearance information. The resulting 3D mesh generated through NeRFMeshing is not only physically accurate but also capable of real-time rendering on various devices. This advancement represents a significant step towards bridging the gap between neural radiance fields and traditional 3D mesh representations, opening up new possibilities for enhanced visualization and simulation in computer graphics applications.