Gaussian Splatting in Style(Related Work)

3D Scene representation

Complex 3D scenes can be represented in many ways. One of the most common and well known ways is that of using point clouds. Similar to point clouds, some works use a voxel grid for representing a 3D scene.

Voxel grids are not the most ideal choice for scene representation due to their high memory footprint.

Neural implicit scene based methods, led by some works making use of signed distance fields(SDFs) and occupancies respectively, helped solve this memory issue.

In principle, implicit neural representations can generate meshes and surfaces upto an arbitrarily high resolution.

Recently, NeuralAngelo set a new benchmark as it could recover a highly detailed 3D surface of a large-scale scene from a set of input images.

Following a similar direction neural radiance fields, also known as NeRFs[50] are also an implicit method.

---

Novel View Synthesis

The task of generating unseen views of an object/scene given a collection of input images is known as novel view synthesis.

Classical works aimed at generating novel views directly from the given set of multi-view ima