This thesis describes and evaluates Dense Surface Models (DSMs), a new technique for building point distribution models of surfaces, from raw input data. DSMs can be used on data from a wide range of surface acquisition systems without preprocessing since they do not require that the surfaces be closed or even locally manifold, and can cope well with holes and spikes in the surfaces. This is an advantage over comparable techniques, which impose such constraints on the input. The core of the DSM algorithm is as follows. Adense correspondence is made between the surfaces using thin-plate spline warping guided by means of a small set of hand-placed landmarks. The area of interest is automatically defined by a threshold on a measure of the closeness of the correspondence at each point. Apoint distribution model is then built using the vertices from the trimmed and densely-corresponded surfaces. The key benefit of using models of the whole surface is illustrated by the large improvement in classification on face shape that is obtained when using DSMs as compared

Facial images present important visual variations, due to several parameters. Here, we focus on the identity and expression parts. We study the hypothesis of (image-based) automatic separability of identity from expressions. Indeed, sign language speakers using videos need a tool able to offer anonymity to their sign productions and such a tool has to modify the part of the facial image carrying identity features without degrading expressive part, needed for comprehension. We present here models of the face space and how they can be used to anonymize facial images. 1

Abstract not found

Abstract not found