The diverse implementations of 3D Gaussian Splatting (3DGS) lead to varied user experiences concerning quality, speed, and operational requirements. This study intends to analyze these variations by categorizing them from several perspectives. Our categorization will specifically highlight innovations in initialization (such as LiDAR), training (including multi-scale methods), and rendering (encompassing platform compatibility and optimization). Through this survey, we provide a comprehensive guide to the rapidly evolving field of real-time 3D reconstruction.
Sogo et al. (Wed,) studied this question.