FaceVerse

FaceVerse数据集是由清华大学和蚂蚁集团联合创建的，专注于东亚人脸的细粒度3D模型。该数据集融合了大规模的RGB-D图像（60,000张）和高保真3D头部扫描模型（2,000个），旨在提供高泛化能力和高保真度的面部细节。创建过程中采用了从粗到细的结构，首先从大规模RGB-D图像生成基础参数模型，然后通过高保真扫描模型训练的条件StyleGAN架构来丰富面部几何和纹理细节。该数据集的应用领域包括电影、视频游戏和混合现实等，旨在解决3D人脸模型在准确性、照片级真实细节和可编辑性方面的挑战。

The FaceVerse dataset, co-developed by Tsinghua University and Ant Group, focuses on fine-grained 3D models of East Asian human faces. This dataset integrates a large-scale collection of RGB-D images (60,000 in total) and high-fidelity 3D head scan models (2,000 in total), aiming to deliver facial details with high generalizability and photorealistic fidelity. A coarse-to-fine framework is adopted during its construction: basic parametric models are first generated from the large-scale RGB-D image collection, then a conditional StyleGAN architecture trained on the high-fidelity scan models is employed to enrich facial geometric and texture details. Application scenarios of this dataset cover film, video games, mixed reality and other fields, and it aims to address the challenges faced by 3D human face models in terms of accuracy, photorealistic details and editability.