User:Bralani/GSoC2024/Abstract

Ray tracing involves computationally expensive calculations and geometry with varying degrees of complexity. The idea for this project is to use neural networks during 3D rendering as an acceleration method. This work is intended to build on research AMD presented and later worked on in BRL-CAD. The general idea is to train a neural net on a model that is otherwise expensive to render (e.g., a NURBS model or complex hierarchical model), then use that neural net for ray intersection queries. The hypothesis being explored is whether neural nets can be leveraged for arbitrary shot line queries, accurately and more efficiently than via direct computation.