Nils Einecke, "Analysis of nonlinear dimensionality reduction for view-based object parameterization", 2007.

Compact and descriptive object representation is a key element for object recognition and tracking. Traditional view-based object representations usually consist of a number of alternative view representatives, leading to a spotty representation in a high dimensional feature space. What is lacking in such a representation is the notion that each view representative is related to low dimensional parameters that characterize the objects appearance variation (like scale, rotation, etc.). If there are sufficient view representatives, they form a low dimensional manifold in the high dimensional feature space. Such manifolds are continuously parameterizable, meaning that neighboring representatives will have similar appearance parameters. The idea is that these manifolds can be utilized as parameterizable object representations. Unsupervised, nonlinear dimensionality reduction algorithms can be used to extract the low dimensional, parameterizable manifold. This diploma thesis is divided into two parts. The first part gives an overview of state-of-the-art nonlinear dimensionality reduction approaches. The three most promising approaches are selected and explained in detail. In the second part of this thesis the selected approaches are investigated concerning their ability to generate parameterizable object representations from multiple object views. Emerging problems are discussed and possible solutions are proposed. Furthermore, a first application for this new kind of object representation is proposed and briefly analyzed.