形状约束的三维人脸组件形变模型

传统的三维人脸形变模型是通过对大量的三维人脸数据进行学习，构建描述人脸三维形状和纹理的参数模型，通过模型优化完成对二维人脸图像的三维重构。但是，实际中大量的训练样本是很难获得的，这导致形变模型描述能力的不完善，制约了它的应用。如将整个人脸看成由若干个组件组合而成，则在样本数不变的情况下降低了描述空间的维数，提高了模型的描述能力。但是在重构人脸图像时必须解决组件间三维空间的重叠合并，并且随着组件数目的增加，模型参数也随之增加，所以对优化算法也提出了更高的要求。为了解决形变模型的这些困难，提出了一种全局模型和组件模型的折中算法，即在形状上保持全局约束而纹理上进行组件匹配，从而在算法性能和算法复杂度之间获得了一个有效的平衡。

Morphable model with shape-constrained 3D face components

Traditional 3D face morphable model constructs the space of the shape parameters and texture parameters by studying those from lots of 3D face samples. But the model can not describe the whole morphable space and can not reconstruct 3D shape well from 2D face image. To solve the problem, a face was constructed by some components, such as eyes, mouth, etc. By this method, the morphable model can construct the parameters space better. But, this kind of component model has more parameters and the overlaps between those components are difficult to deal with. In this paper, we used the shape model as a whole and the texture model as components. In this way, we can get a balance between performance and complication.