PDF version of this document

Real-time Active Appearance Models

Some machines are able to deal with small-sized fitting in real-time [,]. It is possible to track faces in a video (frame rate should then typically be 24 frames/second and 15 at the minimum), but the resolution catered for is often relatively low (less than 100x100 pixel). Applications that respond so quickly were made far more practical owing to a multi-resolution (multi-scale) approach (see below). In order to decrease the total run-time, varying increasing image resolutions become available for selection at each search iteration. Finer resolution images are usually used at the later stages of the search, whereas low-resolution (coarse) ones at the very start. Since the similarity between the model and the target is poor ab ovo, the resolution (and hence the scale of the objects) will have little effect on the fitting. Some visual examples of AAM search are shown in [].

An additional advantage that a multi-resolution approach offer is its notable improvement of structures. It even allows fitting to be more robust to large displacements from the target. This is due to its treatment of a large image as if it was a smaller one - one in which structures are represented by a smaller number of pixels.

Figure: A multi-resolution approach illustrated. Coarser representations are shown at the top levels and the original image lies at the bottom.