This talk presents a detailed explanation for all the parts of the Gaussian splatting rasterizer and how the cuda_rasterizer.cu implements the theory behind the Gaussian splatting. Spherical harmonics for view dependent color, frustrum culling for the culling rasterizers in camera space, splatting the Gaussians from 3D to 2D screen space, rendering the pixles, tiling on GPU to improve performance, the covariance matrix and at the end most important of all, the backward pass that enables optimizing algorithms suck as Adam or SGD to optimize the Gaussians with respect to a given input image and produce novel views of the scene are discussed.
No slides are available for this talk.