Gaussian Adaptive Density Control with SDF Constraints (GADC-SDF) for 3D Human Avatar Modeling

Three-dimensional Gaussian Splatting (3DGS) methods have achieved real-time rendering and high-precision modeling in 3D human avatar modeling. However, most existing methods perform Gaussian sphere cloning, splitting, and pruning without explicit geometric constraints in density control, resulting in inadequate detail reconstruction. To address this, we propose a Gaussian Adaptive Density Control method with SDF constraints (GADC-SDF). Specifically, we introduce the Signed Distance Field (SDF) as explicit geometric constraints for adaptive density control in 3DGS optimization. First, SDF voxel grids are constructed from the SMPL-X mesh. Then, SDF values are applied to prune spheres outside the human body, while additional Gaussian spheres are cloned around the human surface. Furthermore, SDF gradients are utilized to split more Gaussian spheres in detail-rich regions. Experiments conducted on X-Humans, UPB, and ZJU_MoCap datasets demonstrate that our method matches the performance of state-of-the-art baselines across most quantitative metrics while providing a modest improvement in perceptual quality.