Utilizing periodic feature-enhanced neural-field modeling for the photorealistic representation of human head avatars

This paper introduces a groundbreaking neural representation technique known as periodic feature-enhanced neural-field modeling (PNM) tailored for 3D models. PNM has been meticulously crafted to proficiently capture the intricate surface geometry of 4D human avatars, yielding a multitude of applications across various domains, including remote conferencing, livestream marketing, short video webcasting, VR/AR/XR applications, and the video game and movie industry. While current neural modeling approaches excel in representing either low-frequency or high-frequency surface details, they often fall short when simultaneously addressing both aspects, leading to subpar overall model quality. To overcome this limitation, PNM leverages position encoding and periodic activation, harnessing its Fourier properties to deliver enhanced representation of high-frequency data, all the while preserving smooth, noise-free surfaces. Our experiments substantiate PNM’s superiority, surpassing state-of-the-art methods in terms of both quantitative and qualitative model reconstruction quality and the portrayal of high-frequency geometry details. Finally, we apply PNM in the digitization pipeline for 4D human avatars and Metaverse applications, demonstrating its remarkable visual performance in dynamic scenarios.