Adaptive Levels of Detail for Human Gaussian Splats with Hierarchical Embedding

This poster presents a novel method to balance human Gaussian models' visual quality and computational costs through hierarchical Gaussian embedding. Leveraging the Gaussian Splatting technique, our method constructs human Gaussian models with varying levels of detail (LoD) by embedding Gaussians on the mesh model. Our method comprises three steps: model preparation and Gaussian initialization, Gaussian Splatting optimization, and iterative LoD generation. Evaluation against conventional Gaussian Splatting demonstrates comparable image fidelity metrics across varying observation distances, highlighting superior performance in terms of computational costs. Our method optimizes computational efficiency by controlling the number of Gaussians while preserving visual quality. This hierarchical embedding method presents an avenue for achieving highly realistic and low computational cost human Gaussian models, paving the way for multiple human Gaussian models or expansive Gaussian scenes.