Implicit
The scene is encoded as a function that can be queried at any 3D point, for example a neural network. NeRF predicts color and density, an SDF predicts the distance to the surface.
Explicit
The scene is stored as discrete elements: points, triangles, voxels or 3D Gaussians.
| Implicit | Explicit | |
|---|---|---|
| Examples | NeRF, SDF, occupancy field | point cloud, mesh, voxels, 3D Gaussians |
| Storage | network weights or function | list of elements |
| From the lecture | NeRF: editing is hard, training takes time | 3DGS: extremely fast rendering |
Intuition
After NeRF (2020), the field moved back towards explicit representations with 3D Gaussian Splatting (2023), mainly because of rendering speed.
Appears in
- Lecture 1, History: NeRF (implicit) and 3D Gaussian Splatting (explicit)
- Lecture 1, Gaussian Splatting: “fully explicit”, unlike NeRF
- Lecture 5, Representations: pros and cons, inside/outside vs. sampling
- Lecture 6.1, Representations: voxels, point clouds, meshes, implicits
- Lecture 6.1, Meshes vs. implicits: control vs. topology and details
- Lecture 7, NeRF vs. 3DGS: implicit vs. explicit, same compositing
- Lecture 8, Summary: 3DGS as a fully explicit representation