| （英） |
We introduce a novel method for high-fidelity free-viewpoint human synthesis from sparse camera views. Unlike prior approaches requiring dense inputs or scene-specific optimization, our method handles self-occlusion, complex poses, and non-rigid deformations (e.g., clothing, hair) under limited views. Using deep learning model with geometry and appearance-aware design, our method preserves fine surface details and ensures consistency across viewpoints. Experiments on the THuman dataset show superior performance over 3D Gaussian Splatting-based baselines in both quantitative metrics and visual quality. This approach is applicable to practical applications such as VR, telepresence, and immersive communication. |