End-to-end high-fidelity avatar reconstruction using 3D Gaussian splatting with monocular video as the sole input

With the rapid development of Augmented Reality (AR) and Virtual Reality (VR), 3D avatar reconstruction has garnered extensive attention from researchers. However, challenges remain in multi-view-based avatar reconstruction, particularly in accurately aligning features such as hair, clothing, and facial details. Additionally, limitations in avatar representation can lead to unrealistic outcomes under varying lighting conditions. This paper addresses these challenges by introducing a novel tile-based rasterization method that enhances the 3D Gaussian Splatting (3DGS) framework for avatar reconstruction. Our approach utilizes Gaussian ellipsoids initialized through Patch Match Stereo (PMS) and employs both active and passive optimization strategies to refine Gaussian internal parameters and perform effective density filtering. This innovation results in high-fidelity reconstructions without visual artifacts, eliminating the need for facial parameter models and background sampling guidance. Furthermore, our method enables the generation of avatar radiation fields using only monocular video input, simplifying the reconstruction process and reducing execution costs. Experimental results demonstrate that our method outperforms traditional techniques and existing avatar radiation field methods, showcasing its flexibility and lightweight pipeline architecture. This work achieved a competitive results in avatar rendering performance metrics, providing a streamlined solution for high-quality avatar representation.