基于矩阵费雪分布的三维人脸变形模型

3D Face Morphable Model Based on Matrix Fisher Distribution

三维人脸的精确表示有利于各种计算机视觉和图形应用。然而,由于数据离散化和模型线性化,在目前的研究中获取准确的身份和表情线索仍然具有挑战性。本文提出了一种新的三维可变形人脸模型,来学习具有隐式神经表示的非线性连续空间。它构建了两个明确的解纠缠变形场来分别建模与身份和表情相关联的复杂形状,并且引入了一个神经混合场,自适应地混合一系列局部场来学习复杂的细节。其次,我们发现姿态参数在网络中可以更好地被解纠缠,对于人脸变形过程中发生的姿态变换,我们利用基于旋转矩阵的费雪分布矩阵来表示人脸姿态的角度,并模拟头部旋转的不确定性。实验表明我们的方法在人脸细节建模和姿态估计方面具有优越性。

The accurate representation of 3D faces is beneficial to various computer vision and graphics applications.However,due to data discretization and model linearity,it is still challenging to obtain accurate identity and expression cues in current research.In this paper,we propose a new 3D deformable face model to learn a nonlinear continuous space with implicit neural representations.It constructs two explicit disentanglement deformation fields to model the complex shapes associated with identity and expression respectively,and introduces a neural hybrid field to learn complex details by adaptively mixing a series of local fields.Secondly,we find that the pose parameters can be better disentangled in the network.For the pose transformation during face deformation,we use the Fisher distribution matrix based on the rotation matrix to represent the angle of the face pose and simulate the uncertainty of the head rotation.Experiments show that our method has advantages in face detail modeling and pose estimation.