Human Avatar Generation Using 3D Gaussian Splatting With a Focus on Non-Rigid Deformations

This paper proposes a method for reconstructing clothed humans based on 3D Gaussian Splatting (3DGS). Applying static 3DGS to the dynamic problem of human reconstruction is challenging due to the complexity of non-rigid deformations and detailed representation of clothing. To address these issues, we optimized 3DGS to align with non-rigid transformations based on the Splatting Avatar technique. As a result, our method enabled the generation of novel view images with better PSNR accuracy compared to the Splatting Avatar method.