Neural Implicit Representation in Modern Computer Graphic

Recently, neural implicit representations gained popularity in computer graphics as an effective approach to modelling and rendering of various 3D scenes. Contrary to typical representations like polygon meshes, volumetric voxels or point clouds, neural implicit representations define scene geometry, its visual appearance and properties using continuous function approximated with a neural network. In this paper, we consider some of the most popular neural implicit representations used nowadays, which include Neural Radiance Fields (NeRF), Deep Signed Distance Functions (DeepSDF), SIREN networks, Gaussian splatting, Instant Neural Graphics Primitives, and NeRF extensions. All of these neural implicit methods enable novel view synthesis and efficient 3D reconstruction from images. The analysis will include an exploration of what particular technique allows better quality, reduced rendering time, better scalability or dealing with such real-life challenges as dynamic scenes, reflective surfaces or rendering large outdoor environments. At the same time, the paper will consider other related ideas like hybrid representations in Plenoxels, using tensor decompositions or neural implicit representations combined with voxel grids. By comparing different scientific papers dedicated to neural implicit representations, it is possible to identify their advantages and disadvantages as well as the ways to overcome possible difficulties in the future.