PDF version of this document
The strengths of active appearance models have often been demonstrated using facial images. Faces are a challenging type of data to cope with due to wide variation in gesture, inter-subject differences, facial hair, gender, lighting conditions and age which are some of the more dominant factors. It means that a good solution to face interpretation and recognition can lead to progress in other aspects of analysis, e.g. industrial inspection, medical data analysis, etc.
Commercial interest in this area has been another motive for increased experimentation with faces. Access control and gesture recognition are amongst the many possible uses of systems that investigate images (or sequences of images) of human appearance. The innate characteristics of the large number of different faces to be considered makes this problem far from trivial and it remains an excellent evaluation tool in the field of computer vision and not only model-based image analysis. Another frequently-used benchmarking data type is handwritten digits.
One common use of AAM's is for medical image analysis. In order to visualise shape and perform some measurements there needs to be a flexible model that handles anatomical variability and change, for example the expansion of an organ after some period of time. Contrariwise, some objects are not expected to show great variability, but abnormalities need to be detected. Industrial inspection products are an example of systems domain that follows such guidelines and they can often rely on edge detection and point-to-point measurements. These are simpler in the sense that they are quicker in operation and they resemble mammalian vision. Human eyesight is said to be highly sensitive to edges (c/f Mach banding). Sonka et al. [11] provide many more details on those topics.