Mixed 3D Gaussian for Dynamic Scenes Representation and Rendering

Junyuan Guo, Teng Wang, Chao Wang

Abstract: High-quality reconstruction and rendering of 3D scenes has been a long-standing problem in CV and CG. Most previous methods rely on implicit representations, which require significant time overhead and limit practical applications. Several explicit representations were proposed to reduce time overhead later, but none of them can achieve high-quality real-time rendering. 3D Gaussian Splatting achieves outstanding rendering quality and speed, but it cannot be applied to dynamic scenes. In this paper, we propose a new dynamic representation called mixed Gaussian. Specifically, we use separate Gaussian models to represent the dynamics and statics of the scene. For the dynamic part, we introduce a new deformation field to model the motion of points and extend the dynamic gaussian to hyper space for dealing with complex topological changes in the scene. Experiments show that our approach achieves superior rendering quality and real-time rendering compared to previous methods.

External IDs:dblp:conf/icmcs/GuoWW24