Depth-Supervised and Curvature-Optimized 3D Gaussian Splatting for Scene Reconstruction
Year: 2025
Type: article
Abstract: Recently, 3DGS has achieved revolutionary advancements in dynamic rendering and novel view synthesis tasks. The Adaptive Density Control (ADC) algorithm plays a pivotal role in 3DGS, utilizing an image gradient-based average gradient threshold to regulate the splitting and cloning of Gaussians and optimize the density of Gaussian distributions. However, relying solely on an average gradient threshold to adjust Gaussians may lead to some Gaussians exhibiting erroneous opacity or neglecting the contributions of Gaussians in regions with smaller gradients to the rendering process. To overcome these limitations, we introduce depth as a geometric constraint, which adjusts the scale of Gaussians based on depth to ensure more coherent frequency information. Furthermore, after the ADC algorithm, we optimize the Gaussians in dense regions by incorporating curvature-based assessments, ensuring that all Gaussians maintain proper occlusion relationships in the final rendering. We conducted extensive tests on datasets such as synthetic Blender and Mip-NeRF360, and our method demonstrated superior performance compared to 3DGS and other state-of-the-art approaches.
Source: IEEE Access
Authors: Antong Li, Lutao Wang, Ziwei Wang
Institution: Chengdu University of Information Technology
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W4411270838