Image Coding with Active Appearance Models

Abstract

Image coding is the task of representing a set of images as accurately as possible using a fixed number of parameters. One well known example is the linear coding problem that leads to Principal Components Analysis (PCA). Although optimal in a certain sense, PCA has limited coding power. A large number of parameters are often required to code a set of images accurately. In this paper we consider image coding using Active Appearance Models (AAMs). AAMs are a class of generative non-linear models (although linear in both shape and appearance) which have received a great deal of recent attention in the computer vision literature. We describe what it means to code and decode with AAMs, pose the optimal coding problem, and propose an algorithm to solve it. Our algorithm can also be interpreted as an unsupervised model building algorithm.