Baking Neural Radiance Fields for Real-time View Synthesis
Published:
goal: run nerf on a macbook
- train a nerf -> store grid values at a sparse grid
- opacity, diffuse color, feature vector for view dependent affects
- feature vector is integrated, and then transformed by an mlp -> rendered per ray
- this mlp can be trained after storing -> fine tuning
- Cauthy loss for density: sparse prior
- on coarse samples only