We propose a model of appearance and a matching method which combines `global' models (in which a few parameters control global appearance) with local elastic or optical-flow-based methods, in which deformation is described by many local parameters together with some regularisation constraints. We use an Active Appearance Model (AAM) as the global model, which can match a statistical model of appearance to a new image rapidly. However, the amount of variation allowed is constrained by the modes of the model, which may be too restrictive (for instance when insufficient training examples are available, or the number of modes is deliberately truncated for efficiency or memory conservation). To compensate for this, after global AAM convergence, we allow further local model deformation, driven by local AAMs around each model node. This is analogous to optical flow or ‘demon’ methods of non-linear image registration. We describe the technique in detail, and demonstrate that allowing this extra freedom can improve the accuracy of object location with only a modest increase in search time. We show the combined method is more accurate than either pure local or pure global model search.
CITATION STYLE
Cootes, T. F., & Taylor, C. J. (2000). Combining elastic and statistical models of appearance variation. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 1842, pp. 149–163). Springer Verlag. https://doi.org/10.1007/3-540-45054-8_10
Mendeley helps you to discover research relevant for your work.