PDF version of this document
Shape and appearance models are often used to represent data (visually) whose properties are roughly known in advance. These statistical models are defined to be flexible enough to generalise to different types of legal data examples, yet to preserve some invariants and constraints so that no illegal examples are judged to be acceptable. The way in which such models work is that they deform to fit a feature in a given image and the allowable variance is dictated by some values that are incorporated into the model. These values are derived from what is called a training set which defines legal sets of values. This process is the analogue of ``teaching'' a system how to make sensible decisions. Subsequently, some performance evaluation of the trained system is required to infer its accuracy, for example an unbiased error rate. Much more detail can be found in [1].
The use of such models has been quite successful, but accuracy and speed are still two hot topics. The search for good models continues as more demanding applications of higher resolution and higher dimensionality become available. Another problem worth solving is the automatic annotation of images. Not only can it save valuable time of experts in a field, but it can also help in the acquisition of a large number of reliable, precise, unbiased and inexpensive annotated images. With more data handled without human intervention, more input is available to train classification systems such as shape and appearance models.
Section 2 explains some of the key concepts that later on clarify how active appearance models work. It also introduces some concepts that can aid in solving the existing problems and deficiencies of active appearance models as described in Section 3 onwards. More details are available in the referenced material and only short explanatory notes are provided to keep this document sufficiently broad in scope. The last section summarises the issues and concludes on the measures to be taken to tackle them. No substantial developments or ideas are proposed in this paper as it strives to just provide a general overview and a roundup of the state of existing ASM/AAM ``technology''.