Flexible Models of Human Faces for the Analysis and Synthesis of Images
Thomas R. Vetter · MPG.PuRe (Max Planck Society) · 1999
When only a single image of a face is available, can we generate new images of the face across changes in viewpoint or illumination? The approach presented in this chapter acquires its knowledge about possible image changes from other faces and transfers this prior knowledge to a novel face image. Such learning strategies are well known by humans. In contrast, for image synthesis or image analysis little is known of how such a knowledge could be automatically acquired or how such a system could be implemented. In recent years we have developed the concept of linear object classes and implemented an application for human faces [1, 2, 3]. The method allows us to compute novel views of a face from a single image. The method draws, on the one hand, on a general flexible face model that is learned automatically from examples and, on the other hand, on an algorithm that allows for matching this flexible model to a novel face image. In an analysis by synthesis loop the novel image is reconstructed by the model. The novel image now can be described or coded through the internal model parameters that are necessary to reconstruct the image. The design of the model allows also for synthesizing new views of the face.