Model-Based Matching by Linear  Combinations of Prototypes (Michael Jones and Tomaso Poggio)

1. Drawbacks/Weaknesses

There are several drawbacks to this approach,
First is that the level of user interaction is fairly high.  The method of determining the pixelwise correspondence works much more successfully when aided by user interaction (setting points, etc.).  It would be nice if the process were more automated,  since most applications of this process (security, object verification, etc.) are generally automated systems.

 Second is the restrictions on similarity of the novel image to the prototypes.  Due to the process of warping the prototypes  the reference prototype and the error minimization, the novel image cannot have any transformations within itself, i.e. the image must contain the same basic structure as the database images to be successfully matched.

2. Future Directions

It would be useful to re-work the algorithm to handle a more robust area of matching, i.e.  to handle varied structures in either the novel image or the prototypes.

Also a level of automation would be beneficial to the acceptability of this process.

3.  Project Plans

I plan on trying to implement a version of this process, and to test its robustness under various conditions (varied transformations, different shapes, etc.), also to determine any enhancements that could be added to it.