Fourier Feature Activated Neural-field Modelling For Human Avatar Representation

In this paper, we introduce a novel neural representation method called Fourier Feature Activated Neural-field Modelling (FFANM) for 3D models. FFANM is designed to accurately and efficiently capture the surface geometry of 4D human avatars, which has numerous applications in the fields of remote conferencing, livestreaming marketing, short video web-cast, VR/AR, and video games and movie industry. While existing neural modelling methods are capable of representing either low-frequency or high-frequency surface geometry, they cannot do so simultaneously, resulting in poor overall representation quality. To address this limitation, we propose FFANM, which incorporates position encoding and periodic activation to leverage its Fourier properties for better representation of high-frequency information, while maintaining smooth surfaces free from noise. Our experiments demonstrate that FFANM outperforms state-of-the-art methods both quantitatively and qualitatively in terms of overall model reconstruction quality and high-frequency geometry details representation. Finally, We apply our proposed method, FFANM, in the 4D human avatar digitization pipeline and Metaverse application, and its superior visual performance demonstrates its capability in dynamic scenarios.