Embodied AI Glossary中文Nerfstudio / gsplat
Nerfstudio / gsplat（NeRF 与高斯泼溅开源工具库）AdvancedOpen-source toolkits for reconstructing 3D scenes from photos, covering both NeRF and Gaussian splatting.
Nerfstudio is an open-source neural radiance field (NeRF — representing a 3D scene with a neural network) framework from a UC Berkeley team, presented at SIGGRAPH 2023, providing data processing, training, and a web viewer, with methods like Nerfacto built in. gsplat is a CUDA-accelerated 3D Gaussian splatting (representing a scene as a large number of colored 3D Gaussian points) rendering and training library from the same team, used by projects such as Nerfstudio's own Splatfacto. In robotics, they're commonly used to scan a real scene into a renderable 3D asset for real-to-sim transfer, Gaussian-splatting-based simulation, and data augmentation.
ExampleFilm a video circling a tabletop with a phone, process and train it through Nerfstudio, and get a Gaussian-splatting scene that can be rendered from novel viewpoints.
- Related
- Neural Radiance Fields (NeRF) · 3D Gaussian Splatting · Real-to-Sim · Gaussian-Splatting-based Simulation · COLMAP · Novel View Synthesis
- Sources
- Nerfstudio 官网 (Chinese)
gsplat GitHub