Title
3D Gaussian Splatting for Real-Time Radiance Field Rendering
(SIGGRAPH 2023, Kerbl & Kopanas et al.)
Abstract
Radiance Field
- NN -> costly to train & render
- Recent faster methods -> trade off speed for quality
- No real-time display rate for unbounded scene + 1080p rendering
Proposed Method
- State-of-the-art visual quality
- Competitive training time
- High-quality real-time novel-view synthesis at 1080p
3 key elements
- Represent the scene with 3D Gaussians from sparse points
- They preserve desirable properties of continuous volumetric radiance fields for scene optimization
- Avoid computation in empty space
- Interleaved optimization/density control of 3D Gaussians
- optimizing anisotropic covariance to achieve an accurate representation of the scene
- Fast visibility-aware rendering algorithm that supports anisotropic splatting
1. Introduction
2 ways of 3D representations
Meshes and Points
- most common 3D scene representations
- explicit & good fit for GPU/CUDA based rasterization
NeRF
- continuous scene representations
- typically optimizing a MLP using volumetric ray-marching
- Most efficient radiance field -> interpolating values stored in voxel/hash grids or points
- Stochastic sampling required for rendering -> costly and can result in noise.
3D Gaussian
- Combines best of both worlds.