PTHUMAN3D: 3D Gaussian Human Avatar Modeling with the Poincaré Ball and the Triplane Representation

3D human avatar modeling is the key technology in the field of digital entertainment. However, existing methods still face significant challenges in handling complex human poses and recovering fine geometric details. In order to solve the above problems, a 3D Gaussian realistic human avatar modeling method based on the Poincaré ball and the triplane representation is proposed. Our method utilizes 3D Gaussian Splatting (3DGS) to model both 3D geometry and appearance. In our method, the triplane representation is introduced to generate the geometry features of the 3D human body, and the proposed hyperbolic U-net based on the Poincaré ball is used to generate the features of the UV maps. The shape decoder then uses the UV features and the geometry features to obtain the rendering parameters. The renderer uses the obtained rendering parameters to generate the 3D human avatar. Moreover, we design the hyperbolic perceptual loss to optimize the network. Experiments on several challenging datasets demonstrate the advantages of our method over the existing approaches with regard to both visual quality and the representative evaluating criteria.