Overview of 3D Gaussian Splatting

3D Gaussian Splatting is a rasterization technology that can render photorealistic scenes learned from a small number of image samples in real time.

Unlike traditional graphics' triangle rasterization, it represents scenes using Gaussians, with each Gaussian defined by Position, Covariance, Color, and Alpha parameters.

The main operating steps are as follows:

• SfM (Structure from Motion): Estimates a 3D point cloud from a set of images.
• Gaussian Conversion: Converts the estimated points into Gaussians.
• Training: Optimizes the Gaussian parameters using Stochastic Gradient Descent (SGD). During this process, Gaussians are split or pruned to refine detail.
• Differentiable Rasterization: Performs fast and efficient rendering by projecting Gaussians into 2D, sorting them by depth, and blending pixels.