Anti-Aliasing for Multiscale 3DGS

In this work, an Anti-Aliasing technique for Visual Simultaneous Localization and Mapping (VSLAM) with 3D Gaussian Splatting (3DGS) is formulated. We will examine how to use the frequency information to construct a LoD hierarchy without prior knowledge of the scene. This goal will be achieved by optimizing a set of Gaussians in multiple spatial frequencies on the fly in a coarse-to-fine manner.

Existing Level of Detail techniques that enable optimizing and rendering Large-Scale Scenes with 3DGS depend on sparse point clouds and known camera poses as initialization. Recent near real-time approaches do not rely on solid initialization but use a continuous image feed to obtain camera poses and initialize Gaussians from an estimate of the monocular depth.

This work aims to connect these two contrary approaches and to enable higher-quality reconstruction when sampling the set of Gaussians at lower rates than used in the training images.