Instant-NGP (Instant Neural Graphics Primitives) is a method that uses a multi-resolution hash encoding to train and evaluate neural graphics representations such as neural radiance fields far faster than earlier approaches.
Semantic Classification
Content
- Instant-NGP (Instant Neural Graphics Primitives) is a technique introduced by NVIDIA that accelerates neural graphics primitives, including neural radiance fields, by replacing large coordinate networks with a small network combined with a trainable multi-resolution hash encoding of spatial features. This reduces training of a radiance field from hours to seconds or minutes on a single GPU.
- The hash encoding stores feature vectors in compact hash tables indexed by position at several resolutions, letting a small multilayer perceptron interpret them. The approach is used for neural radiance fields, signed distance functions and image fitting, and it underpins fast capture-to-render pipelines for telepresence and 3D reconstruction.