Trường bức xạ thần kinh

NeRF, introduced in the 2020 research paper by Mildenhall and colleagues, represents a static scene with a neural function used for view synthesis. Its conceptual input is a three-dimensional position and a viewing direction with two angular degrees of freedom. The implementation can express that direction as a three-component unit vector, so the conceptual five dimensions should not be confused with the number of stored scalar coordinates in every implementation. The network produces volume density and RGB appearance. Density is not simply one finished pixel’s opacity: rendering depends on density accumulated along a ray and on the distance intervals sampled. The original architecture constrains density to depend on position, while allowing color to depend on both position and viewing direction. That separation supports multiview consistency and direction-dependent appearance, without establishing physical correctness everywhere. Rendering samples the field along a camera ray and composites contributions through a volume-rendering formulation, including transmittance from earlier parts of the ray. Training compares rendered views with observed images and optimizes the representation. The original procedure requires the training camera poses; a set of photographs without a defined pose-estimation or joint-learning method is not the same setup. Positional encoding supplies functions of coordinates that help the evaluated MLP represent high-frequency detail. Original per-scene optimization and rendering costs motivated later approaches. Instant-NGP, for example, studies multiresolution hash encoding and a smaller network under its own configurations. These results should be compared with the actual representation, hardware, and task rather than called a universal hours-to-seconds conversion.