Transfer Discriminative Logmaps

Abstract

In recent years, transfer learning has attracted much attention in multimedia. In this paper, we propose an efficient transfer dimensionality reduction algorithm called transfer discriminative Logmaps (TDL). TDL finds a common feature so that 1) the quadratic distance between the distribution of the training set and that of the testing set is minimized and 2) specific knowledge of the training samples can be conveniently delivered to or shared with the testing samples. Drawing on this common feature in the representation space, our objective is to develop a linear subspace in which discriminative and geometric information can be exploited. TDL adopts the margin maximization to identify discriminative information between different classes, while Logmaps is used to preserve the local-global geodesic distance as well as the direction information. Experiments carried out on both synthetic and real-word image datasets show the effectiveness of TDL for cross-domain face recognition and web image annotation.