Chapter 4 - 3D Morphable face model, a unified approach for analysis and synthesis of images

Abstract

This chapter focuses on the 3D morphable model (3DMM), a unified approach for the analysis and synthesis of images. This model uses the 3D representation and the correspondence principle to account for any individual, viewed from any angles and under any light direction. The chapter discusses the parameters of variation in images of human faces, two- or three-dimensional image models, image analysis by model fitting, and a morphable face model. It also compares fitting algorithms including point-distribution model and active-shape model, inverse compositional image alignment algorithm, 2d+3d active-appearance model, linear shape and texture fitting algorithm. The second component of an analysis-by-synthesis loop is the model inversion, or analyzing algorithm. The most general and accurate algorithm proposed so far is the stochastic Newton optimization (SNO), whose update at each interaction is based on the first and second derivatives of a MAP energy function. The first derivatives are computed at each interaction, thereby, favoring accuracy over efficiency. A stochastic optimization scheme reduces the risk of locking into a local minimum. Additionally to SNO that gives greater importance to accuracy and generality, there are other fitting algorithms that favor efficiency at the expense of the domain of application and the precision. Using 3DMM, the principles of the major fitting algorithms are outlined; and their advantages and predicaments described. The algorithms reviewed are based on active-shape model (ASM), active-appearance model (AAM), and inverse compositional image alignment (ICIA); ICIA applied to 3DMM, 2D+3D AAM, and the linear shape and texture fitting algorithm.