Classification of Dissimilarity Data via Sparse Representation

摘要

Pairwise dissimilarity representations are common practice in several applications of computer vision, since they constitute a powerful alternative to the traditional vectorial representations. During the previous decades many techniques aiming to quantify the dissimilarity between two objects were developed. The problem is that most of these measures tend to produce non-Euclidean and/or non-metric dissimilarity data, although they seem to perform quite well on a variety of tasks. Recently it has been shown that non-Euclidean properties of dissimilarity data may contain useful and discriminative information. In this context, classical embedding of dissimilarity data into a vector space, can imply information loss. An alternative option is the representation into the dissimilarity space, where each object is represented by it's dissimilarity to a set of prototypes. Such a space has the mathematical properties which allow the incorporation of more advanced classifiers, beyond the Nearest Neighbour and the k-NN which are usually the case. In the current work we aim to combine the flexibility of dissimilarity representations with the discriminative ability of the well-established sparse representation-based classification scheme (Wright, 2010), in order to enhance the classification performance on dissimilarity data. The proposed DS-SRC framework has been evaluated on three datasets, derived from different computer vision tasks. The results demonstrate the ability of DS-SRC to improve the classification accuracy, regardless of the special characteristics of each dataset.