Conclusions and Future Work

This chapter will summarize the results obtained in the thesis, point out possible applications and directions for future research. The chapters will be treated in order of apparence if possible (some topics aaect more than one chapter). The introduction treats model-based coding, which is a technique to code image sequences at very low rates. The focus is set on face sequences, where two unre-solved problems are identiied. The texture extraction problem is the problem of acquiring the rst texture. In order to do this, the model must rst be adapted to the face in the image. The second problem is that the number of bits needed to represent the texture is several orders of magnitude greater than that for the motion information of a frame. This causes unacceptable delay for the rst image and makes it impossible to update the texture due to, e.g., changes in lighting. Chapter 2 commences by regarding an image as a point in image space. It is shown that the set of all face images, the face image region, F, is a non-convex subregion of image space. A geometrical normalization mapping, GN, is introduced in order to solve the non-convexity problem, and experiments strongly suggest that GN(F) is indeed convex. Points in GN(F) (i.e., normalized face images, called textures) can thus be well described with linear basis vectors. To nd appropriate basis vectors, a principal component analysis is performed on