Visual Object Categorization using Distance-Based Discriminant Analysis

Abstract

Abstract

Iterative Majorization Approach to the Distance-based Discriminant Analysis

“…Hence, the goal of this analysis was to assess the effect of applying a DDA transformation on the accuracy of the NN classifier. The error rates of NN and DDA+NN data classification experiments are presented in Table 1, showing a consistent improvement Kosinov (2003) for details) using the ETH80 database also revealed the importance of the length constraint, introduced in Section 3 to avoid overfitting. The results of these tests demonstrated up to 20% better classification accuracy for the lengthconstrained version of the method.…”

“…Our preliminary studies, Kosinov (2003), have shown that neither straightforward gradient descent nor some of the state-of-the-art optimization routines are suitable for solving the above optimization problem mostly due to susceptibility to local minima, adverse dependence on the initial value, and difficulties related to the discontinuities of the derivative of (3). However, by deriving some approximations of S W (T ) and S B (T ) one can make the task of minimizing log J(T ) criterion amenable to a simple iterative procedure based on the majorization method (Borg andGroenen (1997), de Leeuw (1977), Heiser (1995)), which we discuss in the following section.…”

“…where p ij is equal to −1 and −1/ d B ij (T ) 2 for C and G, respectively (see Kosinov (2003) for derivation details and a description of an alternative faster method based on Taylor series expansion). Finally, combining results ( 6) and ( 8), we obtain a majorizing function of the log J(T ) optimization criterion:…”

Hierarchical ensemble learning for multimedia categorization and autoannotation

“…boat, sailboat, ferryboat, rowboat, at the expense of precision. Using the DDA baseline classifiers [8,9] for each concept C i ∈ H, the following precision and recall results on the test set vocabulary were obtained (see Figure 3). As seen from the figure, the naturally high recall results boosted by keyword group retrieval, Figure 3(a) do not necessarily correspond to high frequency common concepts emphasizing the importance of the concept co-occurrence factors, while the significantly lower precision values for complex concepts, such as church, fence, boat, Figure 3(b), indicate that these words are much more often retrieved as a group of semanticallyrelated keywords, rather than individually.…”