BibTeX entry

@PHDTHESIS{201102Kirill_Sidorov,
  AUTHOR={Kirill Sidorov},
  TITLE={Groupwise Non-Rigid Registration For Automatic Construction
    Of Appearance Models Of The Human Craniofacial Complex For
    Analysis, Synthesis, And Simulation},
  SCHOOL={Cardiff University},
  MONTH=Feb,
  YEAR=2011,
  URL={http://www.bmva.org/theses/2011/2011-sidorov.pdf},
}

Abstract

In this thesis, the problem of automatic construction of statistical appearance models from examples is considered. The key step in statistical appearance modelling is establishing spatial correspondences between examples in order that statistics on the corresponding features may be computed. This is known as registration. Groupwise registration methods, which aim to consider useful information from the entire ensemble at once when searching for correspondences, have been shown in the literature to be superior to pairwise methods. However, the groupwise approach to registration is generally computationally expensive due to the large dimensionality of the search space in which the globally optimal solution is searched. A novel, fast and reliable, stochastic algorithm is proposed to solve the problem of groupwise non-rigid registration of large ensembles of images quickly and more accurately than state of the art methods. The efficiency of the proposed approach stems from novel dimensionality reduction techniques specific to the problem of groupwise image registration and from comparative insensitivity of the adopted optimisation scheme (Simultaneous Perturbation Stochastic Approximation (SPSA)) to the high dimensionality of the search space. The proposed image registration algorithm is then generalised to the case of textured 3D surfaces, allowing groupwise non-rigid registration of 3D data, such as produced by widely available 3D surface scanners. In evaluation of these approaches we show a high robustness and success rate, fast convergence on various types of test data, including facial images featuring large degrees of both inter- and intra-person variation, and show considerable improvement in terms of accuracy of solution and speed compared to traditional methods. Finally, a novel application of 3D appearance modelling is proposed: a faster than real-time algorithm for statistically constrained quasi-mechanical simulation. Experiments demonstrate superior realism, achieved in the proposed method by employing statistical appearance models to drive the simulation, in comparison with the comparable state of the art quasi-mechanical approaches.