A principle contribution of this research is the further advancement of a model of visual attention that exploits distinctiveness as a criterion for visual interest. The advantage of this model is that it precludes any explicit or implicit domain-dependent assumptions about landmarks. Based on this model, we developed a formal definition of a landmark and formulated and implemented a method for their extraction. The extraction method was shown to be fast and reliable.
In order to characterise the variation in appearance and image geometry of landmarks, we proposed a method for visual tracking which matches landmarks based on appearance and applies no assumptions or constraints on real-world geometry, or on the pose of the camera. Note that while our experiments were all conducted at a fixed orientation, this was performed in order to constrain the dimensionality of the configuration-space, as opposed to its geometry. The tracking method proved robust in a variety of scenes, only demonstrating minor degradation as the sample spacing grew larger, mostly due to aspects of self-similarity, and large changes in view from one viewpoint to its nearest neighbour.