A 66.6 FPS High Quality Gaussian Splats Rendering FPGA Processor with Reconfigurable Computation Architecture
Hongseok Lee, Gwangtae Park, Won-Hoon Park, Wooyoung Jo, Jongjun Park, Hoi‐Jun Yoo · 2024
Recently, 3D Gaussian Splatting (3DGS) has been a novel solution in rendering 3D scene geometry. Unlike prior works, such as Neural Radiance Fields (NeRF), 3DGS can represent scenes using anisotropic covariance parameters as in Fig. 1 [1], enabling high quality rendering in both transparent and reflective scenes with reduced training data [2]. Key stages of rendering process is as follows:1) Frustum Culling - removing gaussians out of view bounds, 2) $2 D$ Feature Computation - computing image parameters from 3D scene parameters, 3) Sorting - aligning gaussians by depth order, 4) Volume Rendering (VR) - accumulating weighted sum of sorted Gaussians to obtain pixel colors.