Large-scale cross-media retrieval by Heterogeneous Feature Augmentation

Media types in heterogeneous source are usually represented in different dimensions; this makes cross-media retrieval hard to process. In this paper, we utilize a new domain adaptation to solve Heterogeneous domain adaptation (HDA) problem in cross-media retrieval using Heterogeneous Feature Augmentation (HFA). First, different dimensions of features are transformed into a common subspace by learning an intermediate variable, and augmented the transformed data with their original features and ones; second, in retrieval stage, we compute the similarity and rank the query results by bag-based reranking method. Experiments on two real-world large-scale image datasets and one text document dataset were conducted; we set two search tasks in the experiment, one is from image to text, and the other is from text to image, the experiment results demonstrate the superiority of our method compared with several newly proposed cross-media retrieval methods.