Zhejiang University Research Team Introduces New AI Framework Human101: Single-View Video Reconstruction for 3D Digital Humans
-
In the field of virtual reality, quickly creating customized virtual characters and enabling interaction with them is a common application scenario. However, previous methods often struggled to balance between rapid reconstruction of high-fidelity 3D digital humans and real-time rendering. In this research, Human101 is proposed, a novel framework capable of generating high-fidelity dynamic 3D human models from just a single-view video with a training time of 100 seconds, and rendering at speeds exceeding 100 FPS.
Unlike previous NeRF-based pipelines, Human101 ingeniously applies a human-centric forward Gaussian animation method by adjusting 3D Gaussian parameters to improve rendering speed. Experiments show that this method achieves significant improvements in frames per second while delivering comparable or superior rendering quality.
To address the slow rendering and convergence issues in traditional methods, Human101 adopts a more practical and improved approach. First, it employs a rasterization rendering pipeline instead of the traditional voxel rendering method to enhance rendering speed. Second, it selects a more easily optimized representation method, namely 3D Gaussians, thereby reducing optimization time to just a few minutes.
Key features of Human101 include:
Human101通过引入创新的技术和实用的方法，成功地解决了单视角视频中快速建模和实时渲染之间的矛盾，为实时交互应用和沉浸式虚拟现实体验提供了可行的解决方案。