Abstract
3D Gaussian Splatting (3DGS) renders pixels by rasterizing Gaussianprimitives, where the rendering resolution and the primitive number, concludedas the optimization complexity, dominate the time cost in primitiveoptimization. In this paper, we propose DashGaussian, a scheduling scheme overthe optimization complexity of 3DGS that strips redundant complexity toaccelerate 3DGS optimization. Specifically, we formulate 3DGS optimization asprogressively fitting 3DGS to higher levels of frequency components in thetraining views, and propose a dynamic rendering resolution scheme that largelyreduces the optimization complexity based on this formulation. Besides, weargue that a specific rendering resolution should cooperate with a properprimitive number for a better balance between computing redundancy and fittingquality, where we schedule the growth of the primitives to synchronize with therendering resolution. Extensive experiments show that our method acceleratesthe optimization of various 3DGS backbones by 45.7% on average while preservingthe rendering quality.