Realistic 3D Human Avatar Creation for Metaverse

This paper presents a complete process of creating lifelike 3D avatar from a single image for the Metaverse. Firstly, we adopt the Position map Regression Network algorithm (PR-Net) [1], which is a deep learning based CNN model to reconstruct 3D faces in .obj format, composed of a .obj file, .mtl file and .png file from a human image. Due to the highly detailed reconstructed 3D face model, i.e., large number of vertices (43,866) produced by the PRNet which is not directly applicable to Metaverse applications, retopology of the model is required. Therefore, the 3D face model is manually retopologized to reduce the number of vertices down to 239 and obtained the optimized 3D face mesh. Subsequently, the pre-built head model is merged with 3D face mesh to achieve a complete human head. Thenceforth, we performed the texture alignment through UV mapping for the accuracy of texture mapping. We solely used a single 3D face mesh as our standard face mesh for all the reconstructed 3D faces. Our goal is to create realistic 3D human avatars to use in the Metaverse. Therefore, we used the publicly accessible body, hair, shoes, and clothing to complete our 3D human avatars. We also decorate our 3D human avatars with accessories and traditional Thai dress, especially for female avatars, to promote the culture of Thailand.