Automatic learning of appearance face models

Since Principal Component Analysis (PCA) technique has been applied by Sirovich and Kirby to parameterize the face, many computer vision researches have used eigen-whatever techniques to construct linear models of optical flow, shape, graylevel, etc. One drawback of such a technique is the need to learn the model, since it is required to gather aligned data, usually with the tedious and inaccurate hand cropping process. This paper describes a robust algorithm for automatically learning an appearance subspace of objects performing rigid motion through an image sequence, given a manual initialization of the regions of support (masks) in the first frame. The learning process is posed as a continuous optimization problem and it is solved with a mixture of stochastic and deterministic techniques achieving sub-pixel accuracy. Additionally, we learn the dynamics of the motion and appearance parameters for scene characterization and point out the benefits of working with modular eigenspaces (ME). Preliminary results of automatic learning a modular eigenface model with applications to real time video conferencing, human computer interaction and actor animation are reported.