NeRF: a scene as a continuous function
On 19 March 2020 researchers from Berkeley, Google Research and San Diego described NeRF: a fully connected network that takes a point in space and a viewing direction and returns density and colour, and from a set of photographs with known camera poses it synthesises new views of complex scenes.
Why it matters
A three-dimensional scene could now be stored as the weights of one small network and rendered from any side by volume rendering. It opened a line of its own for recovering geometry from photographs.
The input is a five-dimensional coordinate: position x, y, z and the viewing direction. Without positional encoding the network does not reproduce fine detail. The weights take 5 MB, 3,000 times less than LLFF, which needs over 15 GB per scene.