PDF version of this document

Active appearance model search

Finally, using the model above and the correlation recorded for that model, it is possible to carry out the search as introduced in the beginning of this section.

The search is basically reliant on error or similarity measures calculated after each attempted parameterisation. Each such reparameterisation is initially guided by the matrices (images) that express correlating between the modes of appearance change and the intensity values.

The model, as shown in Figure 1, is placed within the image frame, close enough to its target. How close it should be put to the target is an issue that will not to be explained in any real detail, but true convergence may never be reached if bad initialisation takes place. The algorithm will most possibly then terminate when it reaches a local minima.

The basic search algorithm, expressed in a simplified way, is as follows:

• Take appearance model in its current state and do:
  • Compute $\delta I$, i.e. the difference between the model and the image.
  • Re-adjust parameter values $c_{i}$. Use the matrices learned off-line to make good choices.
  • Computer $\delta I$ and save new appearance model if better results are obtained. If not, adjustment according to a coefficient $k=1.5,0.5,0.25$ might be worth assigning to $\delta c_{i}$ in order to achieve better results.
• While not converged or improvements are still being made.

The technique of matching an appearance model to an image is described in greater detail with some examples in [10]. It is also worth mentioning that in practice, in order to decrease the total run-time, varying increasing image resolutions are selected in the search iterations. This technique is a very common one in computer graphics and image interpretation tasks. A pyramid can be used to describe the data available for choice. The figure below shows how the size of the image quadruples (doubles in each of the two axes) at each stage of the pyramid, where the base of this pyramid is level 0 (level numbers increase upwards). Finer resolution images are at the bottom and low-resolution coarse ones at the top. The searching process typically begins at the very top of the pyramid and declares convergence only once it has reached the full resolution that is not lossy, i.e. it captures the whole pre-existent data.

Figure 2: Resolution selection