We have been using Appearance Models, statistical models of the shape and texture of objects in images, for several years. However, to fit such a model to an image directly was very time consuming using normal optimization algorithms. Instead we would usually use an Active Shape Model to locate points on a target quickly, then fit the appearance model to the image given those points.
At the end of 1997 we invented a novel algorithm for iteratively improving the fit of an appearance model to an image. It relies on learning the relationship between parameter displacements and residual image errors, allowing a prediction of the current offset of a model when matching an image. The algorithm allows us to fit full appearance models to images much more rapidly than before. This has been successfully applied to face images  and medical images .
We went on to explore the performance of the algorithm, and that of modified versions. The latter included models which were driven by a subset of all the available pixels and ones in which the residual was used to drive the shape variation alone, rather than the combined appearance model parameters .
One of the advantages of working with the company Kestra in 1997 was that I learnt more about object oriented design and the usefulness of design patterns for software. I have spent some time this year re-writing my software libraries to incorporate these new techniques. I have also been able to put the Active Appearance Models and Active Shape Models into the same framework. This will allow direct comparison between the algorithms more easily, and should speed up development time.
The paper describing the AAM algorithm  was a joint winner of the Olympus `Science for Life' Prize for the best paper at the European Conference of Computer Vision, 1998. In addition  won the prize for the best paper at the International Conference on Automatic Face and Gesture Recognition 1999.