Sparse Gaussian Splatting: Dynamic Spatial Sparsification Gaussian Splatting Based on Hash Multi - Resolution Grid

3D Gaussian splatting has brought a new technological revolution to novel view synthesis, achieving real-time rendering, high-quality rendering quality, and fast model training. However, 3DGS has a phenomenon of memory waste in verification, which makes it difficult to be applied on small embedded platforms such as UAV. To address this critical issue, we analyze the main causes of this problem from two perspectives of primitive parameters and primitive distribution and propose three solutions: limiting primitive generation, redundant primitive elimination, and spherical harmonic function pruning. First, we design a constraint function in the covariance scale space of Gaussian primitives. Through large gradients and shape constraints, it affects the shape and spacing of primitive generation, accelerates the iteration speed and reduces the appearance of needle-like artifact primitives. Secondly, during the iteration process, we perform fast multi-resolution hash grid coding on the overall model and calculates the cross-redundancy of each primitive, and adaptively eliminates primitives with high redundancy. Finally, based on the 3DGS rendering function of variable spherical harmonics, we validate the isotropy and high-order spherical harmonic order of the visible surface of primitives and performs adaptive pruning of spherical harmonic parameters. Incorporating all the schemes, we finally achieve a reduction of 70% in the total memory size of the test data set model with almost no drop in model quality, and successfully realizes fast rendering of scenes on mobile devices.