Gaussian Splatting: 3D Reconstruction and Novel View Synthesis, a Review(3)

ABSTRACT

Image-based 3D reconstruction is a challenging task that involves inferring the 3D shape of an object or scene from a set of input images. Learning-based methods have gained attention for their ability to directly estimate 3D shapes. This review paper focuses on SOTA techniques for 3D reconstruction, including the generation of novel, unseen views.

DISCUSSION

Traditionally, 3D scenes have been represented using meshes and points due to their explicit nature and compatibility with rapid GPU/CUDA-based rasterization.

However, recent advancements like NeRF methods settles on for continuous scene representations, employing techniques such as multi layerd perceptron optimization through volumetric ray-marching for novel view synthesis. While continuous representations aid optimization, the stochastic sampling necessary for rendering introduces costly noise.

Gaussian Splatting bridges this gap by leveraging a 3D Gaussian representation for optimization, achieving SOTA visual quality and competitive training times.

Additionally, a tile-based splatting solution ensures real-time rendering with top-tier quality. Gaussian Splatting has delivered some of best results in term of quality and efficiency while rendering 3D scenes.

---

Gaussian Splatting has evolved to handle dynamic and deformable objects by modifying its original representation. This involves incorporating parameters like 3D position, rotation, scaling factor, and spherical harmonics coefficients for color and opacity.

Recent progress in this domain includes the introduction