Query specific re-ranking for improved cross-modal retrieval

Cross-modal retrieval tasks like image-to-text, audio-to-image retrieval, etc. are an important area of research. Different algorithms have been developed to address these tasks. In this work, we propose a novel query specific re-ranking based approach to improve the retrieval performance of any given baseline approach. For each query, the top K-retrieved results of the baseline algorithm are used to compute its class-rank order feature. Based on this feature of the query and the highly relevant examples within the top K-retrieved results, each training example is given a score indicating its relevance to the query, which is finally used to train the query-specific regressor. The new score given by this regressor to each retrieved example is then used to re-rank them. The proposed approach does not require knowledge of the baseline algorithm, and also does not extract additional features from the data. Thus it can be used as an add-on to any existing algorithm for improved retrieval performance. Experiments with several state-of-the-art cross-modal algorithms across different datasets show the effectiveness of the proposed re-ranking algorithm. (C) 2017ElsevierB. V. Allrightsreserved.