Can We Extract 3D Biometrics from 2D Images for Facial Beauty Analysis?
Can We Extract 3D Biometrics from 2D Images for Facial Beauty Analysis?
复制标题
DOI：
10.1145/3436369.3436476
复制
发表时间：
2020-10
期刊：
影响因子：
--
通讯作者：
Wenming Han;Fangmei Chen;Fuming Sun
中科院分区：
文献类型：
--
作者：
Wenming Han;Fangmei Chen;Fuming Sun
几何特征是人脸美容分析中的重要特征，因为它具有清晰的定义和与人类直觉的一致性。已有的工作往往是从受约束环境下拍摄的二维人脸图像中提取这些特征。随着近年来单目重建算法的发展，我们想知道是否有可能从2D图像中重建3D人脸并提取3D生物特征，而不是直接从2D图像中提取特征。如果特征是健壮的，大量的野外人脸图像将可用于面部美容分析。在本文中，我们设计实验来评估2D-3D几何特征的精度和稳健性。以BJUT-3D数据库中的3D人脸作为地面真实点。基于这些3D人脸，我们生成了不同姿势的2D图像，并将它们输入到深度神经网络中，得到重建的3D人脸。分别从地面真实3D人脸、生成的2D人脸和重建的3D人脸中提取比例和角度特征。结果表明，与2D图像相比，2D-3D几何特征对姿态变化具有更强的鲁棒性；单目重建存在模糊性，但误差小于两两个体差异。
Geometric features are important traits in facial beauty analysis due to its clear definition and coherence to our intuition. Existing works often extract these features from 2D face images captured in constrained environment. With the recent development of monocular reconstruction algorithms, we wonder if it is possible to reconstruct 3D faces from 2D images and extract 3D biometrics instead of extracting features from the 2D images directly. If the features are robust, a large amount of in-the-wild face images will be available for facial beauty analysis. In this paper, we design experiments to evaluate the precision and robustness of 2D-3D geometric features. 3D faces in the BJUT-3D database were taken as the ground truth. Based on these 3D faces, we generated 2D images with different poses and fed them to a deep neural network to obtain the reconstructed 3D faces. Ratio and angle features were extracted from the ground truth 3D faces, the generated 2D faces, and the reconstructed 3D faces, respectively. The results show that the 2D-3D geometric features were more robust to pose variations compared with those of 2D images; ambiguities exist in monocular reconstruction, but the error is smaller than the pairwise individual differences.