近日，浙江大学 ReLER 实验室的研究人员提出了一种名为 SIFU 的单图即可重建高质量 3D 人体模型的新方法。这一方法在几何与纹理重建测试中均达到 SOTA，并且在真实世界中具有多种应用场景。
在 AR、VR、3D 打印、场景搭建以及电影制作等多个领域中，高质量的穿着衣服的人体 3D 模型非常重要。然而，传统的创建这些模型的方法需要大量时间，还需要能够捕捉多视角照片的专业设备，此外还依赖于技术熟练的专业人员。相比之下，SIFU 模型能够从单张图像准确重建 3D 人体模型，显著降低成本，并简化独立创作的过程。
SIFU 模型通过在 2D 特征转换到 3D 空间引入人体侧视图作为先验条件，增强几何重建效果。并在纹理优化阶段引入预训练的扩散模型，来解决不可见区域纹理较差的问题。实验证明，SIFU 模型在几何重建与纹理重建中均表现出了最好的效果。
此外，SIFU 模型具有更广阔的应用场景，包括 3D 打印、场景搭建、纹理编辑等。这意味着，该技术有望为相关领域带来更多创新和便利。
英语如下：
Title: Zhejiang University Proposes SIFU Model: Single Image for High-Quality 3D Human Body Reconstruction, Breaching Traditional Methods!
Keywords: 1. SIFU model
Content: Recently, researchers from the ReLER Laboratory at Zhejiang University have proposed a new method called SIFU that can reconstruct high-quality 3D human body models from a single image. This method achieves SOTA in both geometric and texture reconstruction tests and has multiple applications in the real world.
In fields such as AR, VR, 3D printing, scene construction, and film production, high-quality clothing-wearing human 3D models are crucial. However, traditional methods for creating these models require a lot of time, specialized equipment capable of capturing multi-view photos, and skilled professionals. In comparison, the SIFU model accurately reconstructs 3D human body models from a single image, significantly reducing costs and simplifying the process of independent creation.
The SIFU model enhances geometric reconstruction by introducing the prior condition of human side views in transforming 2D features into 3D space. It also addresses the issue of poor texture in invisible areas by introducing a pre-trained diffusion model during the texture optimization phase. The experimental results demonstrate that the SIFU model performs the best in both geometric reconstruction and texture reconstruction.
Furthermore, the SIFU model has broader application scenarios, including 3D printing, scene construction, texture editing, etc. This means that the technology is expected to bring more innovation and convenience to relevant fields.
【来源】https://www.ithome.com/0/746/073.htm
Views: 1