Dynamic density control for 3D Gaussian splatting scene reconstruction

Abstract

Three-dimensional scene reconstruction has broad applications in virtual reality, digital twins, and autonomous driving. Recently, 3D Gaussian Splatting (3DGS) has emerged as a prominent approach in 3D reconstruction due to its favorable balance between reconstruction quality and real-time rendering speed. However, the original 3DGS method employs a fixed Gaussian growth strategy during training, which cannot dynamically adjust the distribution of Gaussians according to the complexity of different scene regions, leading to Gaussian redundancy and insufficient local detail representation. To address this issue, this paper proposes a dynamic density-controlled 3DGS optimization method. The proposed method constructs a Density Complexity Score (DCS) function to evaluate the complexity of local scene regions and dynamically adjusts the number of Gaussian splittings based on the scoring results, thereby allocating more Gaussian resources to complex regions. Experimental results demonstrate that, compared to the original 3DGS method, the proposed approach achieves stable improvements in reconstruction accuracy and rendering efficiency (PSNR improvement of 0.46 dB, approximately 15% reduction in Gaussian count) while maintaining overall rendering quality.