A 2020 method reconstructs photorealistic 3D scenes from a handful of photos. It trains a small network to represent how light travels through a volume, then renders novel viewpoints. The approach transformed computer graphics and 3D capture.
A scene as a function
Representation is implicit. A network maps 3D coordinates and viewing angle to color and density. The scene lives in weights.
Novel views
New angles appear. The model renders viewpoints never photographed. The results are strikingly realistic.
Volume rendering
Physics guides it. Rays are integrated through the learned volume to form pixels. Light transport is modeled directly.
Speed problem
Training was slow. Early versions took hours per scene. Later methods cut this to seconds.
Wide uses
Applications grew. Virtual reality, mapping, and film adopted the technique. Capture became easier.
A successor
Splatting arrived. Explicit point-based methods later rivaled it for speed. The field advanced quickly.
The bottom line
Neural radiance fields reconstruct photorealistic 3D scenes from sparse photos using a small network. They render convincing novel views via volume rendering. They reshaped 3D capture and graphics.
