The Principles of View Invariant, Image-Based Linear Flexible Shape Modelling
Bernard F. Buxton and M. Benjamin Dias
Department of Computer Science, University College London,
Gower Street, London, WC1E 6BT, UK
Flexible shape modelling in which the image of an object is represented by a finite set of landmark points is briefly reviewed. It is argued that variation in the apparent size and shape of the image of an object with change of viewpoint are extrinsic variations. Under weak or para-perspective imaging, such changes should be taken into account by using the affine trifocal tensor to provide a mapping between image triplets. It is shown how the Procrustes error may be extended to accommodate such mappings, how the remaining intrinsic variations may be extracted, a pair of virtual basis views derived, and a statistical model constructed. A hierarchical PCA or coupled view flexible shape model was constructed and shown, on data from a small database of face images, to outperform the coupled view model of Cootes et. al. [10, 8].