Proposed Framework Overview

论文上下文

Paper title: High-Fidelity 3D Gaussian Human Reconstruction via Region-Aware Initialization and Geometric Priors Abstract: Real-time, high-fidelity 3D human reconstruction from RGB images is essential for interactive applications such as virtual reality and gaming, yet remains challenging due to the complex non-rigid deformations of dynamic human bodies. Although 3D Gaussian Splatting enables efficient rendering, existing methods struggle to capture fine geometric details and often produce artifacts such as fused fingers and over-smoothed faces. Moreover, conventional spatial-field-based dynamic modeling faces a trade-off between reconstruction fidelity and GPU memory consumption. To address these issues, we propose a novel 3D Gaussian human reconstruction framework that combines region-aware initialization with rich geometric priors. Specifically, we leverage the expressive SMPL-X model to initialize both 3D Gaussians and skinning weights, providing a robust geometric foundation for precise reconstruction. We further introduce a region-aware density initialization strategy and a geometry-aware multi-scale hash encoding module to improve local detail recovery while maintaining computational efficiency.Experiments on PeopleSnapshot and GalaBasketball show that our method achieves superior reconstruction quality and finer detail preservation under complex motions, while maintaining real-time rendering speed. Passages referencing this figure: \alpha_{j}). (5) Building upon this 3D Gaussian representation, when provided with a target pose as the driving condition, we apply skeletal transformations to spatially deform the canonical Gaussian primitives. Subsequently, these deformed Gaussians are projected onto the 2D imaging plane via 3D Gaussian splatting to synthesize the final rendered image. 3.2 Overall Architecture As illustrated in Figure 1 , the human avatar reconstruction framework proposed in this paper aims to recover high-fidelity, animatable 3D human models from RGB images. Our framework consists of four core stages: human geometric prior extraction, geometry-aware multi-scale hash encoding, canonical space Gaussian modeling, and pose-guided deformation and rendering. By integrating the robust priors of the SMPL-X para osses. The overall loss function is formulated as a direct combination of the L 1 L_{1} loss and a Distance-Structural Similarity Index Measure (D-SSIM) term, L D − S ​ S ​ I ​ M L_{D-SSIM} : L = ( 1 − λ ) ​ L 1 + λ ​ L D − S ​ S ​ I ​ M , L=(1-\lambda)L_{1}+\lambda L_{D-SSIM}, (6) where λ \lambda is empirically set to 0.2, following the optimization configurations of previous works [ 16 , 19 ] . Figure 1: An overview of our proposed framework. 3.3 Region-Aware Density Initialization To facilitate 3D Gaussian deformation based on the Linear Blend Skinning (LBS) algorithm, during our canonical space modeling process, we construct a set of skinned 3D Gaussians and bind them to the corresponding skeleton J J . The formulated skinned 3D Gaussian P s ​ k ​ i ​ n P_{skin} is reconstructed from t