FaceScape

我们提出了一个大规模详细的3D人脸数据集，FaceScape以及相应的基准来评估单视图人脸3D重建。通过对FaceScape数据进行训练，提出了一种新颖的算法来从单个图像输入中预测复杂的可缩放3D人脸模型。FaceScape数据集提供18,760个纹理3D面孔，从938个主题捕获，每个主题有20个特定的表达。3D模型包含孔级面部几何形状，该几何形状也经过处理以在拓扑上均匀。

We propose a large-scale and detailed 3D facial dataset, FaceScape, along with a corresponding benchmark for evaluating single-view facial 3D reconstruction. By training on the FaceScape data, we present a novel algorithm that predicts complex scalable 3D facial models from a single image input. The FaceScape dataset provides 18,760 textured 3D human facial models, captured from 938 subjects, with 20 specific expressions per subject. The 3D models contain pore-level facial geometry, which is also processed to be topologically uniform.

搜集汇总

背景与挑战

背景概述

FaceScape是一个大规模详细的3D人脸数据集，包含18,760个纹理3D面孔，来自938个主题，每个主题有20种特定表情，用于评估单视图人脸3D重建。该数据集支持训练算法从单个图像预测复杂的可缩放3D人脸模型，具有孔级面部几何形状和均匀拓扑处理。由南京大学和肯塔基大学于2020年发布，旨在推动人脸3D建模研究。